• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Budget Fair Queueing (BFQ) I/O scheduler.
4  *
5  * Based on ideas and code from CFQ:
6  * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7  *
8  * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9  *		      Paolo Valente <paolo.valente@unimore.it>
10  *
11  * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
12  *                    Arianna Avanzini <avanzini@google.com>
13  *
14  * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15  *
16  * BFQ is a proportional-share I/O scheduler, with some extra
17  * low-latency capabilities. BFQ also supports full hierarchical
18  * scheduling through cgroups. Next paragraphs provide an introduction
19  * on BFQ inner workings. Details on BFQ benefits, usage and
20  * limitations can be found in Documentation/block/bfq-iosched.rst.
21  *
22  * BFQ is a proportional-share storage-I/O scheduling algorithm based
23  * on the slice-by-slice service scheme of CFQ. But BFQ assigns
24  * budgets, measured in number of sectors, to processes instead of
25  * time slices. The device is not granted to the in-service process
26  * for a given time slice, but until it has exhausted its assigned
27  * budget. This change from the time to the service domain enables BFQ
28  * to distribute the device throughput among processes as desired,
29  * without any distortion due to throughput fluctuations, or to device
30  * internal queueing. BFQ uses an ad hoc internal scheduler, called
31  * B-WF2Q+, to schedule processes according to their budgets. More
32  * precisely, BFQ schedules queues associated with processes. Each
33  * process/queue is assigned a user-configurable weight, and B-WF2Q+
34  * guarantees that each queue receives a fraction of the throughput
35  * proportional to its weight. Thanks to the accurate policy of
36  * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
37  * processes issuing sequential requests (to boost the throughput),
38  * and yet guarantee a low latency to interactive and soft real-time
39  * applications.
40  *
41  * In particular, to provide these low-latency guarantees, BFQ
42  * explicitly privileges the I/O of two classes of time-sensitive
43  * applications: interactive and soft real-time. In more detail, BFQ
44  * behaves this way if the low_latency parameter is set (default
45  * configuration). This feature enables BFQ to provide applications in
46  * these classes with a very low latency.
47  *
48  * To implement this feature, BFQ constantly tries to detect whether
49  * the I/O requests in a bfq_queue come from an interactive or a soft
50  * real-time application. For brevity, in these cases, the queue is
51  * said to be interactive or soft real-time. In both cases, BFQ
52  * privileges the service of the queue, over that of non-interactive
53  * and non-soft-real-time queues. This privileging is performed,
54  * mainly, by raising the weight of the queue. So, for brevity, we
55  * call just weight-raising periods the time periods during which a
56  * queue is privileged, because deemed interactive or soft real-time.
57  *
58  * The detection of soft real-time queues/applications is described in
59  * detail in the comments on the function
60  * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
61  * interactive queue works as follows: a queue is deemed interactive
62  * if it is constantly non empty only for a limited time interval,
63  * after which it does become empty. The queue may be deemed
64  * interactive again (for a limited time), if it restarts being
65  * constantly non empty, provided that this happens only after the
66  * queue has remained empty for a given minimum idle time.
67  *
68  * By default, BFQ computes automatically the above maximum time
69  * interval, i.e., the time interval after which a constantly
70  * non-empty queue stops being deemed interactive. Since a queue is
71  * weight-raised while it is deemed interactive, this maximum time
72  * interval happens to coincide with the (maximum) duration of the
73  * weight-raising for interactive queues.
74  *
75  * Finally, BFQ also features additional heuristics for
76  * preserving both a low latency and a high throughput on NCQ-capable,
77  * rotational or flash-based devices, and to get the job done quickly
78  * for applications consisting in many I/O-bound processes.
79  *
80  * NOTE: if the main or only goal, with a given device, is to achieve
81  * the maximum-possible throughput at all times, then do switch off
82  * all low-latency heuristics for that device, by setting low_latency
83  * to 0.
84  *
85  * BFQ is described in [1], where also a reference to the initial,
86  * more theoretical paper on BFQ can be found. The interested reader
87  * can find in the latter paper full details on the main algorithm, as
88  * well as formulas of the guarantees and formal proofs of all the
89  * properties.  With respect to the version of BFQ presented in these
90  * papers, this implementation adds a few more heuristics, such as the
91  * ones that guarantee a low latency to interactive and soft real-time
92  * applications, and a hierarchical extension based on H-WF2Q+.
93  *
94  * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
95  * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
96  * with O(log N) complexity derives from the one introduced with EEVDF
97  * in [3].
98  *
99  * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
100  *     Scheduler", Proceedings of the First Workshop on Mobile System
101  *     Technologies (MST-2015), May 2015.
102  *     http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
103  *
104  * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105  *     Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
106  *     Oct 1997.
107  *
108  * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
109  *
110  * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
111  *     First: A Flexible and Accurate Mechanism for Proportional Share
112  *     Resource Allocation", technical report.
113  *
114  * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
115  */
116 #include <linux/module.h>
117 #include <linux/slab.h>
118 #include <linux/blkdev.h>
119 #include <linux/cgroup.h>
120 #include <linux/ktime.h>
121 #include <linux/rbtree.h>
122 #include <linux/ioprio.h>
123 #include <linux/sbitmap.h>
124 #include <linux/delay.h>
125 #include <linux/backing-dev.h>
126 
127 #include <trace/events/block.h>
128 
129 #include "elevator.h"
130 #include "blk.h"
131 #include "blk-mq.h"
132 #include "blk-mq-tag.h"
133 #include "blk-mq-sched.h"
134 #include "bfq-iosched.h"
135 #include "blk-wbt.h"
136 
137 #define BFQ_BFQQ_FNS(name)						\
138 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq)			\
139 {									\
140 	__set_bit(BFQQF_##name, &(bfqq)->flags);			\
141 }									\
142 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq)			\
143 {									\
144 	__clear_bit(BFQQF_##name, &(bfqq)->flags);		\
145 }									\
146 int bfq_bfqq_##name(const struct bfq_queue *bfqq)			\
147 {									\
148 	return test_bit(BFQQF_##name, &(bfqq)->flags);		\
149 }
150 
151 BFQ_BFQQ_FNS(just_created);
152 BFQ_BFQQ_FNS(busy);
153 BFQ_BFQQ_FNS(wait_request);
154 BFQ_BFQQ_FNS(non_blocking_wait_rq);
155 BFQ_BFQQ_FNS(fifo_expire);
156 BFQ_BFQQ_FNS(has_short_ttime);
157 BFQ_BFQQ_FNS(sync);
158 BFQ_BFQQ_FNS(IO_bound);
159 BFQ_BFQQ_FNS(in_large_burst);
160 BFQ_BFQQ_FNS(coop);
161 BFQ_BFQQ_FNS(split_coop);
162 BFQ_BFQQ_FNS(softrt_update);
163 #undef BFQ_BFQQ_FNS						\
164 
165 /* Expiration time of async (0) and sync (1) requests, in ns. */
166 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
167 
168 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
169 static const int bfq_back_max = 16 * 1024;
170 
171 /* Penalty of a backwards seek, in number of sectors. */
172 static const int bfq_back_penalty = 2;
173 
174 /* Idling period duration, in ns. */
175 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
176 
177 /* Minimum number of assigned budgets for which stats are safe to compute. */
178 static const int bfq_stats_min_budgets = 194;
179 
180 /* Default maximum budget values, in sectors and number of requests. */
181 static const int bfq_default_max_budget = 16 * 1024;
182 
183 /*
184  * When a sync request is dispatched, the queue that contains that
185  * request, and all the ancestor entities of that queue, are charged
186  * with the number of sectors of the request. In contrast, if the
187  * request is async, then the queue and its ancestor entities are
188  * charged with the number of sectors of the request, multiplied by
189  * the factor below. This throttles the bandwidth for async I/O,
190  * w.r.t. to sync I/O, and it is done to counter the tendency of async
191  * writes to steal I/O throughput to reads.
192  *
193  * The current value of this parameter is the result of a tuning with
194  * several hardware and software configurations. We tried to find the
195  * lowest value for which writes do not cause noticeable problems to
196  * reads. In fact, the lower this parameter, the stabler I/O control,
197  * in the following respect.  The lower this parameter is, the less
198  * the bandwidth enjoyed by a group decreases
199  * - when the group does writes, w.r.t. to when it does reads;
200  * - when other groups do reads, w.r.t. to when they do writes.
201  */
202 static const int bfq_async_charge_factor = 3;
203 
204 /* Default timeout values, in jiffies, approximating CFQ defaults. */
205 const int bfq_timeout = HZ / 8;
206 
207 /*
208  * Time limit for merging (see comments in bfq_setup_cooperator). Set
209  * to the slowest value that, in our tests, proved to be effective in
210  * removing false positives, while not causing true positives to miss
211  * queue merging.
212  *
213  * As can be deduced from the low time limit below, queue merging, if
214  * successful, happens at the very beginning of the I/O of the involved
215  * cooperating processes, as a consequence of the arrival of the very
216  * first requests from each cooperator.  After that, there is very
217  * little chance to find cooperators.
218  */
219 static const unsigned long bfq_merge_time_limit = HZ/10;
220 
221 static struct kmem_cache *bfq_pool;
222 
223 /* Below this threshold (in ns), we consider thinktime immediate. */
224 #define BFQ_MIN_TT		(2 * NSEC_PER_MSEC)
225 
226 /* hw_tag detection: parallel requests threshold and min samples needed. */
227 #define BFQ_HW_QUEUE_THRESHOLD	3
228 #define BFQ_HW_QUEUE_SAMPLES	32
229 
230 #define BFQQ_SEEK_THR		(sector_t)(8 * 100)
231 #define BFQQ_SECT_THR_NONROT	(sector_t)(2 * 32)
232 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
233 	(get_sdist(last_pos, rq) >			\
234 	 BFQQ_SEEK_THR &&				\
235 	 (!blk_queue_nonrot(bfqd->queue) ||		\
236 	  blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
237 #define BFQQ_CLOSE_THR		(sector_t)(8 * 1024)
238 #define BFQQ_SEEKY(bfqq)	(hweight32(bfqq->seek_history) > 19)
239 /*
240  * Sync random I/O is likely to be confused with soft real-time I/O,
241  * because it is characterized by limited throughput and apparently
242  * isochronous arrival pattern. To avoid false positives, queues
243  * containing only random (seeky) I/O are prevented from being tagged
244  * as soft real-time.
245  */
246 #define BFQQ_TOTALLY_SEEKY(bfqq)	(bfqq->seek_history == -1)
247 
248 /* Min number of samples required to perform peak-rate update */
249 #define BFQ_RATE_MIN_SAMPLES	32
250 /* Min observation time interval required to perform a peak-rate update (ns) */
251 #define BFQ_RATE_MIN_INTERVAL	(300*NSEC_PER_MSEC)
252 /* Target observation time interval for a peak-rate update (ns) */
253 #define BFQ_RATE_REF_INTERVAL	NSEC_PER_SEC
254 
255 /*
256  * Shift used for peak-rate fixed precision calculations.
257  * With
258  * - the current shift: 16 positions
259  * - the current type used to store rate: u32
260  * - the current unit of measure for rate: [sectors/usec], or, more precisely,
261  *   [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
262  * the range of rates that can be stored is
263  * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
264  * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
265  * [15, 65G] sectors/sec
266  * Which, assuming a sector size of 512B, corresponds to a range of
267  * [7.5K, 33T] B/sec
268  */
269 #define BFQ_RATE_SHIFT		16
270 
271 /*
272  * When configured for computing the duration of the weight-raising
273  * for interactive queues automatically (see the comments at the
274  * beginning of this file), BFQ does it using the following formula:
275  * duration = (ref_rate / r) * ref_wr_duration,
276  * where r is the peak rate of the device, and ref_rate and
277  * ref_wr_duration are two reference parameters.  In particular,
278  * ref_rate is the peak rate of the reference storage device (see
279  * below), and ref_wr_duration is about the maximum time needed, with
280  * BFQ and while reading two files in parallel, to load typical large
281  * applications on the reference device (see the comments on
282  * max_service_from_wr below, for more details on how ref_wr_duration
283  * is obtained).  In practice, the slower/faster the device at hand
284  * is, the more/less it takes to load applications with respect to the
285  * reference device.  Accordingly, the longer/shorter BFQ grants
286  * weight raising to interactive applications.
287  *
288  * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
289  * depending on whether the device is rotational or non-rotational.
290  *
291  * In the following definitions, ref_rate[0] and ref_wr_duration[0]
292  * are the reference values for a rotational device, whereas
293  * ref_rate[1] and ref_wr_duration[1] are the reference values for a
294  * non-rotational device. The reference rates are not the actual peak
295  * rates of the devices used as a reference, but slightly lower
296  * values. The reason for using slightly lower values is that the
297  * peak-rate estimator tends to yield slightly lower values than the
298  * actual peak rate (it can yield the actual peak rate only if there
299  * is only one process doing I/O, and the process does sequential
300  * I/O).
301  *
302  * The reference peak rates are measured in sectors/usec, left-shifted
303  * by BFQ_RATE_SHIFT.
304  */
305 static int ref_rate[2] = {14000, 33000};
306 /*
307  * To improve readability, a conversion function is used to initialize
308  * the following array, which entails that the array can be
309  * initialized only in a function.
310  */
311 static int ref_wr_duration[2];
312 
313 /*
314  * BFQ uses the above-detailed, time-based weight-raising mechanism to
315  * privilege interactive tasks. This mechanism is vulnerable to the
316  * following false positives: I/O-bound applications that will go on
317  * doing I/O for much longer than the duration of weight
318  * raising. These applications have basically no benefit from being
319  * weight-raised at the beginning of their I/O. On the opposite end,
320  * while being weight-raised, these applications
321  * a) unjustly steal throughput to applications that may actually need
322  * low latency;
323  * b) make BFQ uselessly perform device idling; device idling results
324  * in loss of device throughput with most flash-based storage, and may
325  * increase latencies when used purposelessly.
326  *
327  * BFQ tries to reduce these problems, by adopting the following
328  * countermeasure. To introduce this countermeasure, we need first to
329  * finish explaining how the duration of weight-raising for
330  * interactive tasks is computed.
331  *
332  * For a bfq_queue deemed as interactive, the duration of weight
333  * raising is dynamically adjusted, as a function of the estimated
334  * peak rate of the device, so as to be equal to the time needed to
335  * execute the 'largest' interactive task we benchmarked so far. By
336  * largest task, we mean the task for which each involved process has
337  * to do more I/O than for any of the other tasks we benchmarked. This
338  * reference interactive task is the start-up of LibreOffice Writer,
339  * and in this task each process/bfq_queue needs to have at most ~110K
340  * sectors transferred.
341  *
342  * This last piece of information enables BFQ to reduce the actual
343  * duration of weight-raising for at least one class of I/O-bound
344  * applications: those doing sequential or quasi-sequential I/O. An
345  * example is file copy. In fact, once started, the main I/O-bound
346  * processes of these applications usually consume the above 110K
347  * sectors in much less time than the processes of an application that
348  * is starting, because these I/O-bound processes will greedily devote
349  * almost all their CPU cycles only to their target,
350  * throughput-friendly I/O operations. This is even more true if BFQ
351  * happens to be underestimating the device peak rate, and thus
352  * overestimating the duration of weight raising. But, according to
353  * our measurements, once transferred 110K sectors, these processes
354  * have no right to be weight-raised any longer.
355  *
356  * Basing on the last consideration, BFQ ends weight-raising for a
357  * bfq_queue if the latter happens to have received an amount of
358  * service at least equal to the following constant. The constant is
359  * set to slightly more than 110K, to have a minimum safety margin.
360  *
361  * This early ending of weight-raising reduces the amount of time
362  * during which interactive false positives cause the two problems
363  * described at the beginning of these comments.
364  */
365 static const unsigned long max_service_from_wr = 120000;
366 
367 /*
368  * Maximum time between the creation of two queues, for stable merge
369  * to be activated (in ms)
370  */
371 static const unsigned long bfq_activation_stable_merging = 600;
372 /*
373  * Minimum time to be waited before evaluating delayed stable merge (in ms)
374  */
375 static const unsigned long bfq_late_stable_merging = 600;
376 
377 #define RQ_BIC(rq)		((struct bfq_io_cq *)((rq)->elv.priv[0]))
378 #define RQ_BFQQ(rq)		((rq)->elv.priv[1])
379 
bic_to_bfqq(struct bfq_io_cq * bic,bool is_sync)380 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
381 {
382 	return bic->bfqq[is_sync];
383 }
384 
385 static void bfq_put_stable_ref(struct bfq_queue *bfqq);
386 
bic_set_bfqq(struct bfq_io_cq * bic,struct bfq_queue * bfqq,bool is_sync)387 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
388 {
389 	struct bfq_queue *old_bfqq = bic->bfqq[is_sync];
390 
391 	/* Clear bic pointer if bfqq is detached from this bic */
392 	if (old_bfqq && old_bfqq->bic == bic)
393 		old_bfqq->bic = NULL;
394 
395 	/*
396 	 * If bfqq != NULL, then a non-stable queue merge between
397 	 * bic->bfqq and bfqq is happening here. This causes troubles
398 	 * in the following case: bic->bfqq has also been scheduled
399 	 * for a possible stable merge with bic->stable_merge_bfqq,
400 	 * and bic->stable_merge_bfqq == bfqq happens to
401 	 * hold. Troubles occur because bfqq may then undergo a split,
402 	 * thereby becoming eligible for a stable merge. Yet, if
403 	 * bic->stable_merge_bfqq points exactly to bfqq, then bfqq
404 	 * would be stably merged with itself. To avoid this anomaly,
405 	 * we cancel the stable merge if
406 	 * bic->stable_merge_bfqq == bfqq.
407 	 */
408 	bic->bfqq[is_sync] = bfqq;
409 
410 	if (bfqq && bic->stable_merge_bfqq == bfqq) {
411 		/*
412 		 * Actually, these same instructions are executed also
413 		 * in bfq_setup_cooperator, in case of abort or actual
414 		 * execution of a stable merge. We could avoid
415 		 * repeating these instructions there too, but if we
416 		 * did so, we would nest even more complexity in this
417 		 * function.
418 		 */
419 		bfq_put_stable_ref(bic->stable_merge_bfqq);
420 
421 		bic->stable_merge_bfqq = NULL;
422 	}
423 }
424 
bic_to_bfqd(struct bfq_io_cq * bic)425 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
426 {
427 	return bic->icq.q->elevator->elevator_data;
428 }
429 
430 /**
431  * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
432  * @icq: the iocontext queue.
433  */
icq_to_bic(struct io_cq * icq)434 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
435 {
436 	/* bic->icq is the first member, %NULL will convert to %NULL */
437 	return container_of(icq, struct bfq_io_cq, icq);
438 }
439 
440 /**
441  * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
442  * @q: the request queue.
443  */
bfq_bic_lookup(struct request_queue * q)444 static struct bfq_io_cq *bfq_bic_lookup(struct request_queue *q)
445 {
446 	struct bfq_io_cq *icq;
447 	unsigned long flags;
448 
449 	if (!current->io_context)
450 		return NULL;
451 
452 	spin_lock_irqsave(&q->queue_lock, flags);
453 	icq = icq_to_bic(ioc_lookup_icq(q));
454 	spin_unlock_irqrestore(&q->queue_lock, flags);
455 
456 	return icq;
457 }
458 
459 /*
460  * Scheduler run of queue, if there are requests pending and no one in the
461  * driver that will restart queueing.
462  */
bfq_schedule_dispatch(struct bfq_data * bfqd)463 void bfq_schedule_dispatch(struct bfq_data *bfqd)
464 {
465 	lockdep_assert_held(&bfqd->lock);
466 
467 	if (bfqd->queued != 0) {
468 		bfq_log(bfqd, "schedule dispatch");
469 		blk_mq_run_hw_queues(bfqd->queue, true);
470 	}
471 }
472 
473 #define bfq_class_idle(bfqq)	((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
474 
475 #define bfq_sample_valid(samples)	((samples) > 80)
476 
477 /*
478  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
479  * We choose the request that is closer to the head right now.  Distance
480  * behind the head is penalized and only allowed to a certain extent.
481  */
bfq_choose_req(struct bfq_data * bfqd,struct request * rq1,struct request * rq2,sector_t last)482 static struct request *bfq_choose_req(struct bfq_data *bfqd,
483 				      struct request *rq1,
484 				      struct request *rq2,
485 				      sector_t last)
486 {
487 	sector_t s1, s2, d1 = 0, d2 = 0;
488 	unsigned long back_max;
489 #define BFQ_RQ1_WRAP	0x01 /* request 1 wraps */
490 #define BFQ_RQ2_WRAP	0x02 /* request 2 wraps */
491 	unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
492 
493 	if (!rq1 || rq1 == rq2)
494 		return rq2;
495 	if (!rq2)
496 		return rq1;
497 
498 	if (rq_is_sync(rq1) && !rq_is_sync(rq2))
499 		return rq1;
500 	else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
501 		return rq2;
502 	if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
503 		return rq1;
504 	else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
505 		return rq2;
506 
507 	s1 = blk_rq_pos(rq1);
508 	s2 = blk_rq_pos(rq2);
509 
510 	/*
511 	 * By definition, 1KiB is 2 sectors.
512 	 */
513 	back_max = bfqd->bfq_back_max * 2;
514 
515 	/*
516 	 * Strict one way elevator _except_ in the case where we allow
517 	 * short backward seeks which are biased as twice the cost of a
518 	 * similar forward seek.
519 	 */
520 	if (s1 >= last)
521 		d1 = s1 - last;
522 	else if (s1 + back_max >= last)
523 		d1 = (last - s1) * bfqd->bfq_back_penalty;
524 	else
525 		wrap |= BFQ_RQ1_WRAP;
526 
527 	if (s2 >= last)
528 		d2 = s2 - last;
529 	else if (s2 + back_max >= last)
530 		d2 = (last - s2) * bfqd->bfq_back_penalty;
531 	else
532 		wrap |= BFQ_RQ2_WRAP;
533 
534 	/* Found required data */
535 
536 	/*
537 	 * By doing switch() on the bit mask "wrap" we avoid having to
538 	 * check two variables for all permutations: --> faster!
539 	 */
540 	switch (wrap) {
541 	case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
542 		if (d1 < d2)
543 			return rq1;
544 		else if (d2 < d1)
545 			return rq2;
546 
547 		if (s1 >= s2)
548 			return rq1;
549 		else
550 			return rq2;
551 
552 	case BFQ_RQ2_WRAP:
553 		return rq1;
554 	case BFQ_RQ1_WRAP:
555 		return rq2;
556 	case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
557 	default:
558 		/*
559 		 * Since both rqs are wrapped,
560 		 * start with the one that's further behind head
561 		 * (--> only *one* back seek required),
562 		 * since back seek takes more time than forward.
563 		 */
564 		if (s1 <= s2)
565 			return rq1;
566 		else
567 			return rq2;
568 	}
569 }
570 
571 #define BFQ_LIMIT_INLINE_DEPTH 16
572 
573 #ifdef CONFIG_BFQ_GROUP_IOSCHED
bfqq_request_over_limit(struct bfq_queue * bfqq,int limit)574 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
575 {
576 	struct bfq_data *bfqd = bfqq->bfqd;
577 	struct bfq_entity *entity = &bfqq->entity;
578 	struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
579 	struct bfq_entity **entities = inline_entities;
580 	int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
581 	int class_idx = bfqq->ioprio_class - 1;
582 	struct bfq_sched_data *sched_data;
583 	unsigned long wsum;
584 	bool ret = false;
585 
586 	if (!entity->on_st_or_in_serv)
587 		return false;
588 
589 retry:
590 	spin_lock_irq(&bfqd->lock);
591 	/* +1 for bfqq entity, root cgroup not included */
592 	depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
593 	if (depth > alloc_depth) {
594 		spin_unlock_irq(&bfqd->lock);
595 		if (entities != inline_entities)
596 			kfree(entities);
597 		entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
598 		if (!entities)
599 			return false;
600 		alloc_depth = depth;
601 		goto retry;
602 	}
603 
604 	sched_data = entity->sched_data;
605 	/* Gather our ancestors as we need to traverse them in reverse order */
606 	level = 0;
607 	for_each_entity(entity) {
608 		/*
609 		 * If at some level entity is not even active, allow request
610 		 * queueing so that BFQ knows there's work to do and activate
611 		 * entities.
612 		 */
613 		if (!entity->on_st_or_in_serv)
614 			goto out;
615 		/* Uh, more parents than cgroup subsystem thinks? */
616 		if (WARN_ON_ONCE(level >= depth))
617 			break;
618 		entities[level++] = entity;
619 	}
620 	WARN_ON_ONCE(level != depth);
621 	for (level--; level >= 0; level--) {
622 		entity = entities[level];
623 		if (level > 0) {
624 			wsum = bfq_entity_service_tree(entity)->wsum;
625 		} else {
626 			int i;
627 			/*
628 			 * For bfqq itself we take into account service trees
629 			 * of all higher priority classes and multiply their
630 			 * weights so that low prio queue from higher class
631 			 * gets more requests than high prio queue from lower
632 			 * class.
633 			 */
634 			wsum = 0;
635 			for (i = 0; i <= class_idx; i++) {
636 				wsum = wsum * IOPRIO_BE_NR +
637 					sched_data->service_tree[i].wsum;
638 			}
639 		}
640 		if (!wsum)
641 			continue;
642 		limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum);
643 		if (entity->allocated >= limit) {
644 			bfq_log_bfqq(bfqq->bfqd, bfqq,
645 				"too many requests: allocated %d limit %d level %d",
646 				entity->allocated, limit, level);
647 			ret = true;
648 			break;
649 		}
650 	}
651 out:
652 	spin_unlock_irq(&bfqd->lock);
653 	if (entities != inline_entities)
654 		kfree(entities);
655 	return ret;
656 }
657 #else
bfqq_request_over_limit(struct bfq_queue * bfqq,int limit)658 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
659 {
660 	return false;
661 }
662 #endif
663 
664 /*
665  * Async I/O can easily starve sync I/O (both sync reads and sync
666  * writes), by consuming all tags. Similarly, storms of sync writes,
667  * such as those that sync(2) may trigger, can starve sync reads.
668  * Limit depths of async I/O and sync writes so as to counter both
669  * problems.
670  *
671  * Also if a bfq queue or its parent cgroup consume more tags than would be
672  * appropriate for their weight, we trim the available tag depth to 1. This
673  * avoids a situation where one cgroup can starve another cgroup from tags and
674  * thus block service differentiation among cgroups. Note that because the
675  * queue / cgroup already has many requests allocated and queued, this does not
676  * significantly affect service guarantees coming from the BFQ scheduling
677  * algorithm.
678  */
bfq_limit_depth(blk_opf_t opf,struct blk_mq_alloc_data * data)679 static void bfq_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
680 {
681 	struct bfq_data *bfqd = data->q->elevator->elevator_data;
682 	struct bfq_io_cq *bic = bfq_bic_lookup(data->q);
683 	struct bfq_queue *bfqq = bic ? bic_to_bfqq(bic, op_is_sync(opf)) : NULL;
684 	int depth;
685 	unsigned limit = data->q->nr_requests;
686 
687 	/* Sync reads have full depth available */
688 	if (op_is_sync(opf) && !op_is_write(opf)) {
689 		depth = 0;
690 	} else {
691 		depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(opf)];
692 		limit = (limit * depth) >> bfqd->full_depth_shift;
693 	}
694 
695 	/*
696 	 * Does queue (or any parent entity) exceed number of requests that
697 	 * should be available to it? Heavily limit depth so that it cannot
698 	 * consume more available requests and thus starve other entities.
699 	 */
700 	if (bfqq && bfqq_request_over_limit(bfqq, limit))
701 		depth = 1;
702 
703 	bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
704 		__func__, bfqd->wr_busy_queues, op_is_sync(opf), depth);
705 	if (depth)
706 		data->shallow_depth = depth;
707 }
708 
709 static struct bfq_queue *
bfq_rq_pos_tree_lookup(struct bfq_data * bfqd,struct rb_root * root,sector_t sector,struct rb_node ** ret_parent,struct rb_node *** rb_link)710 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
711 		     sector_t sector, struct rb_node **ret_parent,
712 		     struct rb_node ***rb_link)
713 {
714 	struct rb_node **p, *parent;
715 	struct bfq_queue *bfqq = NULL;
716 
717 	parent = NULL;
718 	p = &root->rb_node;
719 	while (*p) {
720 		struct rb_node **n;
721 
722 		parent = *p;
723 		bfqq = rb_entry(parent, struct bfq_queue, pos_node);
724 
725 		/*
726 		 * Sort strictly based on sector. Smallest to the left,
727 		 * largest to the right.
728 		 */
729 		if (sector > blk_rq_pos(bfqq->next_rq))
730 			n = &(*p)->rb_right;
731 		else if (sector < blk_rq_pos(bfqq->next_rq))
732 			n = &(*p)->rb_left;
733 		else
734 			break;
735 		p = n;
736 		bfqq = NULL;
737 	}
738 
739 	*ret_parent = parent;
740 	if (rb_link)
741 		*rb_link = p;
742 
743 	bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
744 		(unsigned long long)sector,
745 		bfqq ? bfqq->pid : 0);
746 
747 	return bfqq;
748 }
749 
bfq_too_late_for_merging(struct bfq_queue * bfqq)750 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
751 {
752 	return bfqq->service_from_backlogged > 0 &&
753 		time_is_before_jiffies(bfqq->first_IO_time +
754 				       bfq_merge_time_limit);
755 }
756 
757 /*
758  * The following function is not marked as __cold because it is
759  * actually cold, but for the same performance goal described in the
760  * comments on the likely() at the beginning of
761  * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
762  * execution time for the case where this function is not invoked, we
763  * had to add an unlikely() in each involved if().
764  */
765 void __cold
bfq_pos_tree_add_move(struct bfq_data * bfqd,struct bfq_queue * bfqq)766 bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
767 {
768 	struct rb_node **p, *parent;
769 	struct bfq_queue *__bfqq;
770 
771 	if (bfqq->pos_root) {
772 		rb_erase(&bfqq->pos_node, bfqq->pos_root);
773 		bfqq->pos_root = NULL;
774 	}
775 
776 	/* oom_bfqq does not participate in queue merging */
777 	if (bfqq == &bfqd->oom_bfqq)
778 		return;
779 
780 	/*
781 	 * bfqq cannot be merged any longer (see comments in
782 	 * bfq_setup_cooperator): no point in adding bfqq into the
783 	 * position tree.
784 	 */
785 	if (bfq_too_late_for_merging(bfqq))
786 		return;
787 
788 	if (bfq_class_idle(bfqq))
789 		return;
790 	if (!bfqq->next_rq)
791 		return;
792 
793 	bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree;
794 	__bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
795 			blk_rq_pos(bfqq->next_rq), &parent, &p);
796 	if (!__bfqq) {
797 		rb_link_node(&bfqq->pos_node, parent, p);
798 		rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
799 	} else
800 		bfqq->pos_root = NULL;
801 }
802 
803 /*
804  * The following function returns false either if every active queue
805  * must receive the same share of the throughput (symmetric scenario),
806  * or, as a special case, if bfqq must receive a share of the
807  * throughput lower than or equal to the share that every other active
808  * queue must receive.  If bfqq does sync I/O, then these are the only
809  * two cases where bfqq happens to be guaranteed its share of the
810  * throughput even if I/O dispatching is not plugged when bfqq remains
811  * temporarily empty (for more details, see the comments in the
812  * function bfq_better_to_idle()). For this reason, the return value
813  * of this function is used to check whether I/O-dispatch plugging can
814  * be avoided.
815  *
816  * The above first case (symmetric scenario) occurs when:
817  * 1) all active queues have the same weight,
818  * 2) all active queues belong to the same I/O-priority class,
819  * 3) all active groups at the same level in the groups tree have the same
820  *    weight,
821  * 4) all active groups at the same level in the groups tree have the same
822  *    number of children.
823  *
824  * Unfortunately, keeping the necessary state for evaluating exactly
825  * the last two symmetry sub-conditions above would be quite complex
826  * and time consuming. Therefore this function evaluates, instead,
827  * only the following stronger three sub-conditions, for which it is
828  * much easier to maintain the needed state:
829  * 1) all active queues have the same weight,
830  * 2) all active queues belong to the same I/O-priority class,
831  * 3) there are no active groups.
832  * In particular, the last condition is always true if hierarchical
833  * support or the cgroups interface are not enabled, thus no state
834  * needs to be maintained in this case.
835  */
bfq_asymmetric_scenario(struct bfq_data * bfqd,struct bfq_queue * bfqq)836 static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
837 				   struct bfq_queue *bfqq)
838 {
839 	bool smallest_weight = bfqq &&
840 		bfqq->weight_counter &&
841 		bfqq->weight_counter ==
842 		container_of(
843 			rb_first_cached(&bfqd->queue_weights_tree),
844 			struct bfq_weight_counter,
845 			weights_node);
846 
847 	/*
848 	 * For queue weights to differ, queue_weights_tree must contain
849 	 * at least two nodes.
850 	 */
851 	bool varied_queue_weights = !smallest_weight &&
852 		!RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
853 		(bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
854 		 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
855 
856 	bool multiple_classes_busy =
857 		(bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
858 		(bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
859 		(bfqd->busy_queues[1] && bfqd->busy_queues[2]);
860 
861 	return varied_queue_weights || multiple_classes_busy
862 #ifdef CONFIG_BFQ_GROUP_IOSCHED
863 	       || bfqd->num_groups_with_pending_reqs > 0
864 #endif
865 		;
866 }
867 
868 /*
869  * If the weight-counter tree passed as input contains no counter for
870  * the weight of the input queue, then add that counter; otherwise just
871  * increment the existing counter.
872  *
873  * Note that weight-counter trees contain few nodes in mostly symmetric
874  * scenarios. For example, if all queues have the same weight, then the
875  * weight-counter tree for the queues may contain at most one node.
876  * This holds even if low_latency is on, because weight-raised queues
877  * are not inserted in the tree.
878  * In most scenarios, the rate at which nodes are created/destroyed
879  * should be low too.
880  */
bfq_weights_tree_add(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct rb_root_cached * root)881 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
882 			  struct rb_root_cached *root)
883 {
884 	struct bfq_entity *entity = &bfqq->entity;
885 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
886 	bool leftmost = true;
887 
888 	/*
889 	 * Do not insert if the queue is already associated with a
890 	 * counter, which happens if:
891 	 *   1) a request arrival has caused the queue to become both
892 	 *      non-weight-raised, and hence change its weight, and
893 	 *      backlogged; in this respect, each of the two events
894 	 *      causes an invocation of this function,
895 	 *   2) this is the invocation of this function caused by the
896 	 *      second event. This second invocation is actually useless,
897 	 *      and we handle this fact by exiting immediately. More
898 	 *      efficient or clearer solutions might possibly be adopted.
899 	 */
900 	if (bfqq->weight_counter)
901 		return;
902 
903 	while (*new) {
904 		struct bfq_weight_counter *__counter = container_of(*new,
905 						struct bfq_weight_counter,
906 						weights_node);
907 		parent = *new;
908 
909 		if (entity->weight == __counter->weight) {
910 			bfqq->weight_counter = __counter;
911 			goto inc_counter;
912 		}
913 		if (entity->weight < __counter->weight)
914 			new = &((*new)->rb_left);
915 		else {
916 			new = &((*new)->rb_right);
917 			leftmost = false;
918 		}
919 	}
920 
921 	bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
922 				       GFP_ATOMIC);
923 
924 	/*
925 	 * In the unlucky event of an allocation failure, we just
926 	 * exit. This will cause the weight of queue to not be
927 	 * considered in bfq_asymmetric_scenario, which, in its turn,
928 	 * causes the scenario to be deemed wrongly symmetric in case
929 	 * bfqq's weight would have been the only weight making the
930 	 * scenario asymmetric.  On the bright side, no unbalance will
931 	 * however occur when bfqq becomes inactive again (the
932 	 * invocation of this function is triggered by an activation
933 	 * of queue).  In fact, bfq_weights_tree_remove does nothing
934 	 * if !bfqq->weight_counter.
935 	 */
936 	if (unlikely(!bfqq->weight_counter))
937 		return;
938 
939 	bfqq->weight_counter->weight = entity->weight;
940 	rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
941 	rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
942 				leftmost);
943 
944 inc_counter:
945 	bfqq->weight_counter->num_active++;
946 	bfqq->ref++;
947 }
948 
949 /*
950  * Decrement the weight counter associated with the queue, and, if the
951  * counter reaches 0, remove the counter from the tree.
952  * See the comments to the function bfq_weights_tree_add() for considerations
953  * about overhead.
954  */
__bfq_weights_tree_remove(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct rb_root_cached * root)955 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
956 			       struct bfq_queue *bfqq,
957 			       struct rb_root_cached *root)
958 {
959 	if (!bfqq->weight_counter)
960 		return;
961 
962 	bfqq->weight_counter->num_active--;
963 	if (bfqq->weight_counter->num_active > 0)
964 		goto reset_entity_pointer;
965 
966 	rb_erase_cached(&bfqq->weight_counter->weights_node, root);
967 	kfree(bfqq->weight_counter);
968 
969 reset_entity_pointer:
970 	bfqq->weight_counter = NULL;
971 	bfq_put_queue(bfqq);
972 }
973 
974 /*
975  * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
976  * of active groups for each queue's inactive parent entity.
977  */
bfq_weights_tree_remove(struct bfq_data * bfqd,struct bfq_queue * bfqq)978 void bfq_weights_tree_remove(struct bfq_data *bfqd,
979 			     struct bfq_queue *bfqq)
980 {
981 	struct bfq_entity *entity = bfqq->entity.parent;
982 
983 	for_each_entity(entity) {
984 		struct bfq_sched_data *sd = entity->my_sched_data;
985 
986 		if (sd->next_in_service || sd->in_service_entity) {
987 			/*
988 			 * entity is still active, because either
989 			 * next_in_service or in_service_entity is not
990 			 * NULL (see the comments on the definition of
991 			 * next_in_service for details on why
992 			 * in_service_entity must be checked too).
993 			 *
994 			 * As a consequence, its parent entities are
995 			 * active as well, and thus this loop must
996 			 * stop here.
997 			 */
998 			break;
999 		}
1000 
1001 		/*
1002 		 * The decrement of num_groups_with_pending_reqs is
1003 		 * not performed immediately upon the deactivation of
1004 		 * entity, but it is delayed to when it also happens
1005 		 * that the first leaf descendant bfqq of entity gets
1006 		 * all its pending requests completed. The following
1007 		 * instructions perform this delayed decrement, if
1008 		 * needed. See the comments on
1009 		 * num_groups_with_pending_reqs for details.
1010 		 */
1011 		if (entity->in_groups_with_pending_reqs) {
1012 			entity->in_groups_with_pending_reqs = false;
1013 			bfqd->num_groups_with_pending_reqs--;
1014 		}
1015 	}
1016 
1017 	/*
1018 	 * Next function is invoked last, because it causes bfqq to be
1019 	 * freed if the following holds: bfqq is not in service and
1020 	 * has no dispatched request. DO NOT use bfqq after the next
1021 	 * function invocation.
1022 	 */
1023 	__bfq_weights_tree_remove(bfqd, bfqq,
1024 				  &bfqd->queue_weights_tree);
1025 }
1026 
1027 /*
1028  * Return expired entry, or NULL to just start from scratch in rbtree.
1029  */
bfq_check_fifo(struct bfq_queue * bfqq,struct request * last)1030 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
1031 				      struct request *last)
1032 {
1033 	struct request *rq;
1034 
1035 	if (bfq_bfqq_fifo_expire(bfqq))
1036 		return NULL;
1037 
1038 	bfq_mark_bfqq_fifo_expire(bfqq);
1039 
1040 	rq = rq_entry_fifo(bfqq->fifo.next);
1041 
1042 	if (rq == last || ktime_get_ns() < rq->fifo_time)
1043 		return NULL;
1044 
1045 	bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
1046 	return rq;
1047 }
1048 
bfq_find_next_rq(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct request * last)1049 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
1050 					struct bfq_queue *bfqq,
1051 					struct request *last)
1052 {
1053 	struct rb_node *rbnext = rb_next(&last->rb_node);
1054 	struct rb_node *rbprev = rb_prev(&last->rb_node);
1055 	struct request *next, *prev = NULL;
1056 
1057 	/* Follow expired path, else get first next available. */
1058 	next = bfq_check_fifo(bfqq, last);
1059 	if (next)
1060 		return next;
1061 
1062 	if (rbprev)
1063 		prev = rb_entry_rq(rbprev);
1064 
1065 	if (rbnext)
1066 		next = rb_entry_rq(rbnext);
1067 	else {
1068 		rbnext = rb_first(&bfqq->sort_list);
1069 		if (rbnext && rbnext != &last->rb_node)
1070 			next = rb_entry_rq(rbnext);
1071 	}
1072 
1073 	return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1074 }
1075 
1076 /* see the definition of bfq_async_charge_factor for details */
bfq_serv_to_charge(struct request * rq,struct bfq_queue * bfqq)1077 static unsigned long bfq_serv_to_charge(struct request *rq,
1078 					struct bfq_queue *bfqq)
1079 {
1080 	if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1081 	    bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1082 		return blk_rq_sectors(rq);
1083 
1084 	return blk_rq_sectors(rq) * bfq_async_charge_factor;
1085 }
1086 
1087 /**
1088  * bfq_updated_next_req - update the queue after a new next_rq selection.
1089  * @bfqd: the device data the queue belongs to.
1090  * @bfqq: the queue to update.
1091  *
1092  * If the first request of a queue changes we make sure that the queue
1093  * has enough budget to serve at least its first request (if the
1094  * request has grown).  We do this because if the queue has not enough
1095  * budget for its first request, it has to go through two dispatch
1096  * rounds to actually get it dispatched.
1097  */
bfq_updated_next_req(struct bfq_data * bfqd,struct bfq_queue * bfqq)1098 static void bfq_updated_next_req(struct bfq_data *bfqd,
1099 				 struct bfq_queue *bfqq)
1100 {
1101 	struct bfq_entity *entity = &bfqq->entity;
1102 	struct request *next_rq = bfqq->next_rq;
1103 	unsigned long new_budget;
1104 
1105 	if (!next_rq)
1106 		return;
1107 
1108 	if (bfqq == bfqd->in_service_queue)
1109 		/*
1110 		 * In order not to break guarantees, budgets cannot be
1111 		 * changed after an entity has been selected.
1112 		 */
1113 		return;
1114 
1115 	new_budget = max_t(unsigned long,
1116 			   max_t(unsigned long, bfqq->max_budget,
1117 				 bfq_serv_to_charge(next_rq, bfqq)),
1118 			   entity->service);
1119 	if (entity->budget != new_budget) {
1120 		entity->budget = new_budget;
1121 		bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1122 					 new_budget);
1123 		bfq_requeue_bfqq(bfqd, bfqq, false);
1124 	}
1125 }
1126 
bfq_wr_duration(struct bfq_data * bfqd)1127 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1128 {
1129 	u64 dur;
1130 
1131 	if (bfqd->bfq_wr_max_time > 0)
1132 		return bfqd->bfq_wr_max_time;
1133 
1134 	dur = bfqd->rate_dur_prod;
1135 	do_div(dur, bfqd->peak_rate);
1136 
1137 	/*
1138 	 * Limit duration between 3 and 25 seconds. The upper limit
1139 	 * has been conservatively set after the following worst case:
1140 	 * on a QEMU/KVM virtual machine
1141 	 * - running in a slow PC
1142 	 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1143 	 * - serving a heavy I/O workload, such as the sequential reading
1144 	 *   of several files
1145 	 * mplayer took 23 seconds to start, if constantly weight-raised.
1146 	 *
1147 	 * As for higher values than that accommodating the above bad
1148 	 * scenario, tests show that higher values would often yield
1149 	 * the opposite of the desired result, i.e., would worsen
1150 	 * responsiveness by allowing non-interactive applications to
1151 	 * preserve weight raising for too long.
1152 	 *
1153 	 * On the other end, lower values than 3 seconds make it
1154 	 * difficult for most interactive tasks to complete their jobs
1155 	 * before weight-raising finishes.
1156 	 */
1157 	return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1158 }
1159 
1160 /* switch back from soft real-time to interactive weight raising */
switch_back_to_interactive_wr(struct bfq_queue * bfqq,struct bfq_data * bfqd)1161 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1162 					  struct bfq_data *bfqd)
1163 {
1164 	bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1165 	bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1166 	bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1167 }
1168 
1169 static void
bfq_bfqq_resume_state(struct bfq_queue * bfqq,struct bfq_data * bfqd,struct bfq_io_cq * bic,bool bfq_already_existing)1170 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1171 		      struct bfq_io_cq *bic, bool bfq_already_existing)
1172 {
1173 	unsigned int old_wr_coeff = 1;
1174 	bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1175 
1176 	if (bic->saved_has_short_ttime)
1177 		bfq_mark_bfqq_has_short_ttime(bfqq);
1178 	else
1179 		bfq_clear_bfqq_has_short_ttime(bfqq);
1180 
1181 	if (bic->saved_IO_bound)
1182 		bfq_mark_bfqq_IO_bound(bfqq);
1183 	else
1184 		bfq_clear_bfqq_IO_bound(bfqq);
1185 
1186 	bfqq->last_serv_time_ns = bic->saved_last_serv_time_ns;
1187 	bfqq->inject_limit = bic->saved_inject_limit;
1188 	bfqq->decrease_time_jif = bic->saved_decrease_time_jif;
1189 
1190 	bfqq->entity.new_weight = bic->saved_weight;
1191 	bfqq->ttime = bic->saved_ttime;
1192 	bfqq->io_start_time = bic->saved_io_start_time;
1193 	bfqq->tot_idle_time = bic->saved_tot_idle_time;
1194 	/*
1195 	 * Restore weight coefficient only if low_latency is on
1196 	 */
1197 	if (bfqd->low_latency) {
1198 		old_wr_coeff = bfqq->wr_coeff;
1199 		bfqq->wr_coeff = bic->saved_wr_coeff;
1200 	}
1201 	bfqq->service_from_wr = bic->saved_service_from_wr;
1202 	bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1203 	bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1204 	bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1205 
1206 	if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1207 	    time_is_before_jiffies(bfqq->last_wr_start_finish +
1208 				   bfqq->wr_cur_max_time))) {
1209 		if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1210 		    !bfq_bfqq_in_large_burst(bfqq) &&
1211 		    time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1212 					     bfq_wr_duration(bfqd))) {
1213 			switch_back_to_interactive_wr(bfqq, bfqd);
1214 		} else {
1215 			bfqq->wr_coeff = 1;
1216 			bfq_log_bfqq(bfqq->bfqd, bfqq,
1217 				     "resume state: switching off wr");
1218 		}
1219 	}
1220 
1221 	/* make sure weight will be updated, however we got here */
1222 	bfqq->entity.prio_changed = 1;
1223 
1224 	if (likely(!busy))
1225 		return;
1226 
1227 	if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1228 		bfqd->wr_busy_queues++;
1229 	else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1230 		bfqd->wr_busy_queues--;
1231 }
1232 
bfqq_process_refs(struct bfq_queue * bfqq)1233 static int bfqq_process_refs(struct bfq_queue *bfqq)
1234 {
1235 	return bfqq->ref - bfqq->entity.allocated -
1236 		bfqq->entity.on_st_or_in_serv -
1237 		(bfqq->weight_counter != NULL) - bfqq->stable_ref;
1238 }
1239 
1240 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
bfq_reset_burst_list(struct bfq_data * bfqd,struct bfq_queue * bfqq)1241 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1242 {
1243 	struct bfq_queue *item;
1244 	struct hlist_node *n;
1245 
1246 	hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1247 		hlist_del_init(&item->burst_list_node);
1248 
1249 	/*
1250 	 * Start the creation of a new burst list only if there is no
1251 	 * active queue. See comments on the conditional invocation of
1252 	 * bfq_handle_burst().
1253 	 */
1254 	if (bfq_tot_busy_queues(bfqd) == 0) {
1255 		hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1256 		bfqd->burst_size = 1;
1257 	} else
1258 		bfqd->burst_size = 0;
1259 
1260 	bfqd->burst_parent_entity = bfqq->entity.parent;
1261 }
1262 
1263 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
bfq_add_to_burst(struct bfq_data * bfqd,struct bfq_queue * bfqq)1264 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1265 {
1266 	/* Increment burst size to take into account also bfqq */
1267 	bfqd->burst_size++;
1268 
1269 	if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1270 		struct bfq_queue *pos, *bfqq_item;
1271 		struct hlist_node *n;
1272 
1273 		/*
1274 		 * Enough queues have been activated shortly after each
1275 		 * other to consider this burst as large.
1276 		 */
1277 		bfqd->large_burst = true;
1278 
1279 		/*
1280 		 * We can now mark all queues in the burst list as
1281 		 * belonging to a large burst.
1282 		 */
1283 		hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1284 				     burst_list_node)
1285 			bfq_mark_bfqq_in_large_burst(bfqq_item);
1286 		bfq_mark_bfqq_in_large_burst(bfqq);
1287 
1288 		/*
1289 		 * From now on, and until the current burst finishes, any
1290 		 * new queue being activated shortly after the last queue
1291 		 * was inserted in the burst can be immediately marked as
1292 		 * belonging to a large burst. So the burst list is not
1293 		 * needed any more. Remove it.
1294 		 */
1295 		hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1296 					  burst_list_node)
1297 			hlist_del_init(&pos->burst_list_node);
1298 	} else /*
1299 		* Burst not yet large: add bfqq to the burst list. Do
1300 		* not increment the ref counter for bfqq, because bfqq
1301 		* is removed from the burst list before freeing bfqq
1302 		* in put_queue.
1303 		*/
1304 		hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1305 }
1306 
1307 /*
1308  * If many queues belonging to the same group happen to be created
1309  * shortly after each other, then the processes associated with these
1310  * queues have typically a common goal. In particular, bursts of queue
1311  * creations are usually caused by services or applications that spawn
1312  * many parallel threads/processes. Examples are systemd during boot,
1313  * or git grep. To help these processes get their job done as soon as
1314  * possible, it is usually better to not grant either weight-raising
1315  * or device idling to their queues, unless these queues must be
1316  * protected from the I/O flowing through other active queues.
1317  *
1318  * In this comment we describe, firstly, the reasons why this fact
1319  * holds, and, secondly, the next function, which implements the main
1320  * steps needed to properly mark these queues so that they can then be
1321  * treated in a different way.
1322  *
1323  * The above services or applications benefit mostly from a high
1324  * throughput: the quicker the requests of the activated queues are
1325  * cumulatively served, the sooner the target job of these queues gets
1326  * completed. As a consequence, weight-raising any of these queues,
1327  * which also implies idling the device for it, is almost always
1328  * counterproductive, unless there are other active queues to isolate
1329  * these new queues from. If there no other active queues, then
1330  * weight-raising these new queues just lowers throughput in most
1331  * cases.
1332  *
1333  * On the other hand, a burst of queue creations may be caused also by
1334  * the start of an application that does not consist of a lot of
1335  * parallel I/O-bound threads. In fact, with a complex application,
1336  * several short processes may need to be executed to start-up the
1337  * application. In this respect, to start an application as quickly as
1338  * possible, the best thing to do is in any case to privilege the I/O
1339  * related to the application with respect to all other
1340  * I/O. Therefore, the best strategy to start as quickly as possible
1341  * an application that causes a burst of queue creations is to
1342  * weight-raise all the queues created during the burst. This is the
1343  * exact opposite of the best strategy for the other type of bursts.
1344  *
1345  * In the end, to take the best action for each of the two cases, the
1346  * two types of bursts need to be distinguished. Fortunately, this
1347  * seems relatively easy, by looking at the sizes of the bursts. In
1348  * particular, we found a threshold such that only bursts with a
1349  * larger size than that threshold are apparently caused by
1350  * services or commands such as systemd or git grep. For brevity,
1351  * hereafter we call just 'large' these bursts. BFQ *does not*
1352  * weight-raise queues whose creation occurs in a large burst. In
1353  * addition, for each of these queues BFQ performs or does not perform
1354  * idling depending on which choice boosts the throughput more. The
1355  * exact choice depends on the device and request pattern at
1356  * hand.
1357  *
1358  * Unfortunately, false positives may occur while an interactive task
1359  * is starting (e.g., an application is being started). The
1360  * consequence is that the queues associated with the task do not
1361  * enjoy weight raising as expected. Fortunately these false positives
1362  * are very rare. They typically occur if some service happens to
1363  * start doing I/O exactly when the interactive task starts.
1364  *
1365  * Turning back to the next function, it is invoked only if there are
1366  * no active queues (apart from active queues that would belong to the
1367  * same, possible burst bfqq would belong to), and it implements all
1368  * the steps needed to detect the occurrence of a large burst and to
1369  * properly mark all the queues belonging to it (so that they can then
1370  * be treated in a different way). This goal is achieved by
1371  * maintaining a "burst list" that holds, temporarily, the queues that
1372  * belong to the burst in progress. The list is then used to mark
1373  * these queues as belonging to a large burst if the burst does become
1374  * large. The main steps are the following.
1375  *
1376  * . when the very first queue is created, the queue is inserted into the
1377  *   list (as it could be the first queue in a possible burst)
1378  *
1379  * . if the current burst has not yet become large, and a queue Q that does
1380  *   not yet belong to the burst is activated shortly after the last time
1381  *   at which a new queue entered the burst list, then the function appends
1382  *   Q to the burst list
1383  *
1384  * . if, as a consequence of the previous step, the burst size reaches
1385  *   the large-burst threshold, then
1386  *
1387  *     . all the queues in the burst list are marked as belonging to a
1388  *       large burst
1389  *
1390  *     . the burst list is deleted; in fact, the burst list already served
1391  *       its purpose (keeping temporarily track of the queues in a burst,
1392  *       so as to be able to mark them as belonging to a large burst in the
1393  *       previous sub-step), and now is not needed any more
1394  *
1395  *     . the device enters a large-burst mode
1396  *
1397  * . if a queue Q that does not belong to the burst is created while
1398  *   the device is in large-burst mode and shortly after the last time
1399  *   at which a queue either entered the burst list or was marked as
1400  *   belonging to the current large burst, then Q is immediately marked
1401  *   as belonging to a large burst.
1402  *
1403  * . if a queue Q that does not belong to the burst is created a while
1404  *   later, i.e., not shortly after, than the last time at which a queue
1405  *   either entered the burst list or was marked as belonging to the
1406  *   current large burst, then the current burst is deemed as finished and:
1407  *
1408  *        . the large-burst mode is reset if set
1409  *
1410  *        . the burst list is emptied
1411  *
1412  *        . Q is inserted in the burst list, as Q may be the first queue
1413  *          in a possible new burst (then the burst list contains just Q
1414  *          after this step).
1415  */
bfq_handle_burst(struct bfq_data * bfqd,struct bfq_queue * bfqq)1416 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1417 {
1418 	/*
1419 	 * If bfqq is already in the burst list or is part of a large
1420 	 * burst, or finally has just been split, then there is
1421 	 * nothing else to do.
1422 	 */
1423 	if (!hlist_unhashed(&bfqq->burst_list_node) ||
1424 	    bfq_bfqq_in_large_burst(bfqq) ||
1425 	    time_is_after_eq_jiffies(bfqq->split_time +
1426 				     msecs_to_jiffies(10)))
1427 		return;
1428 
1429 	/*
1430 	 * If bfqq's creation happens late enough, or bfqq belongs to
1431 	 * a different group than the burst group, then the current
1432 	 * burst is finished, and related data structures must be
1433 	 * reset.
1434 	 *
1435 	 * In this respect, consider the special case where bfqq is
1436 	 * the very first queue created after BFQ is selected for this
1437 	 * device. In this case, last_ins_in_burst and
1438 	 * burst_parent_entity are not yet significant when we get
1439 	 * here. But it is easy to verify that, whether or not the
1440 	 * following condition is true, bfqq will end up being
1441 	 * inserted into the burst list. In particular the list will
1442 	 * happen to contain only bfqq. And this is exactly what has
1443 	 * to happen, as bfqq may be the first queue of the first
1444 	 * burst.
1445 	 */
1446 	if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1447 	    bfqd->bfq_burst_interval) ||
1448 	    bfqq->entity.parent != bfqd->burst_parent_entity) {
1449 		bfqd->large_burst = false;
1450 		bfq_reset_burst_list(bfqd, bfqq);
1451 		goto end;
1452 	}
1453 
1454 	/*
1455 	 * If we get here, then bfqq is being activated shortly after the
1456 	 * last queue. So, if the current burst is also large, we can mark
1457 	 * bfqq as belonging to this large burst immediately.
1458 	 */
1459 	if (bfqd->large_burst) {
1460 		bfq_mark_bfqq_in_large_burst(bfqq);
1461 		goto end;
1462 	}
1463 
1464 	/*
1465 	 * If we get here, then a large-burst state has not yet been
1466 	 * reached, but bfqq is being activated shortly after the last
1467 	 * queue. Then we add bfqq to the burst.
1468 	 */
1469 	bfq_add_to_burst(bfqd, bfqq);
1470 end:
1471 	/*
1472 	 * At this point, bfqq either has been added to the current
1473 	 * burst or has caused the current burst to terminate and a
1474 	 * possible new burst to start. In particular, in the second
1475 	 * case, bfqq has become the first queue in the possible new
1476 	 * burst.  In both cases last_ins_in_burst needs to be moved
1477 	 * forward.
1478 	 */
1479 	bfqd->last_ins_in_burst = jiffies;
1480 }
1481 
bfq_bfqq_budget_left(struct bfq_queue * bfqq)1482 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1483 {
1484 	struct bfq_entity *entity = &bfqq->entity;
1485 
1486 	return entity->budget - entity->service;
1487 }
1488 
1489 /*
1490  * If enough samples have been computed, return the current max budget
1491  * stored in bfqd, which is dynamically updated according to the
1492  * estimated disk peak rate; otherwise return the default max budget
1493  */
bfq_max_budget(struct bfq_data * bfqd)1494 static int bfq_max_budget(struct bfq_data *bfqd)
1495 {
1496 	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1497 		return bfq_default_max_budget;
1498 	else
1499 		return bfqd->bfq_max_budget;
1500 }
1501 
1502 /*
1503  * Return min budget, which is a fraction of the current or default
1504  * max budget (trying with 1/32)
1505  */
bfq_min_budget(struct bfq_data * bfqd)1506 static int bfq_min_budget(struct bfq_data *bfqd)
1507 {
1508 	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1509 		return bfq_default_max_budget / 32;
1510 	else
1511 		return bfqd->bfq_max_budget / 32;
1512 }
1513 
1514 /*
1515  * The next function, invoked after the input queue bfqq switches from
1516  * idle to busy, updates the budget of bfqq. The function also tells
1517  * whether the in-service queue should be expired, by returning
1518  * true. The purpose of expiring the in-service queue is to give bfqq
1519  * the chance to possibly preempt the in-service queue, and the reason
1520  * for preempting the in-service queue is to achieve one of the two
1521  * goals below.
1522  *
1523  * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1524  * expired because it has remained idle. In particular, bfqq may have
1525  * expired for one of the following two reasons:
1526  *
1527  * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1528  *   and did not make it to issue a new request before its last
1529  *   request was served;
1530  *
1531  * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1532  *   a new request before the expiration of the idling-time.
1533  *
1534  * Even if bfqq has expired for one of the above reasons, the process
1535  * associated with the queue may be however issuing requests greedily,
1536  * and thus be sensitive to the bandwidth it receives (bfqq may have
1537  * remained idle for other reasons: CPU high load, bfqq not enjoying
1538  * idling, I/O throttling somewhere in the path from the process to
1539  * the I/O scheduler, ...). But if, after every expiration for one of
1540  * the above two reasons, bfqq has to wait for the service of at least
1541  * one full budget of another queue before being served again, then
1542  * bfqq is likely to get a much lower bandwidth or resource time than
1543  * its reserved ones. To address this issue, two countermeasures need
1544  * to be taken.
1545  *
1546  * First, the budget and the timestamps of bfqq need to be updated in
1547  * a special way on bfqq reactivation: they need to be updated as if
1548  * bfqq did not remain idle and did not expire. In fact, if they are
1549  * computed as if bfqq expired and remained idle until reactivation,
1550  * then the process associated with bfqq is treated as if, instead of
1551  * being greedy, it stopped issuing requests when bfqq remained idle,
1552  * and restarts issuing requests only on this reactivation. In other
1553  * words, the scheduler does not help the process recover the "service
1554  * hole" between bfqq expiration and reactivation. As a consequence,
1555  * the process receives a lower bandwidth than its reserved one. In
1556  * contrast, to recover this hole, the budget must be updated as if
1557  * bfqq was not expired at all before this reactivation, i.e., it must
1558  * be set to the value of the remaining budget when bfqq was
1559  * expired. Along the same line, timestamps need to be assigned the
1560  * value they had the last time bfqq was selected for service, i.e.,
1561  * before last expiration. Thus timestamps need to be back-shifted
1562  * with respect to their normal computation (see [1] for more details
1563  * on this tricky aspect).
1564  *
1565  * Secondly, to allow the process to recover the hole, the in-service
1566  * queue must be expired too, to give bfqq the chance to preempt it
1567  * immediately. In fact, if bfqq has to wait for a full budget of the
1568  * in-service queue to be completed, then it may become impossible to
1569  * let the process recover the hole, even if the back-shifted
1570  * timestamps of bfqq are lower than those of the in-service queue. If
1571  * this happens for most or all of the holes, then the process may not
1572  * receive its reserved bandwidth. In this respect, it is worth noting
1573  * that, being the service of outstanding requests unpreemptible, a
1574  * little fraction of the holes may however be unrecoverable, thereby
1575  * causing a little loss of bandwidth.
1576  *
1577  * The last important point is detecting whether bfqq does need this
1578  * bandwidth recovery. In this respect, the next function deems the
1579  * process associated with bfqq greedy, and thus allows it to recover
1580  * the hole, if: 1) the process is waiting for the arrival of a new
1581  * request (which implies that bfqq expired for one of the above two
1582  * reasons), and 2) such a request has arrived soon. The first
1583  * condition is controlled through the flag non_blocking_wait_rq,
1584  * while the second through the flag arrived_in_time. If both
1585  * conditions hold, then the function computes the budget in the
1586  * above-described special way, and signals that the in-service queue
1587  * should be expired. Timestamp back-shifting is done later in
1588  * __bfq_activate_entity.
1589  *
1590  * 2. Reduce latency. Even if timestamps are not backshifted to let
1591  * the process associated with bfqq recover a service hole, bfqq may
1592  * however happen to have, after being (re)activated, a lower finish
1593  * timestamp than the in-service queue.	 That is, the next budget of
1594  * bfqq may have to be completed before the one of the in-service
1595  * queue. If this is the case, then preempting the in-service queue
1596  * allows this goal to be achieved, apart from the unpreemptible,
1597  * outstanding requests mentioned above.
1598  *
1599  * Unfortunately, regardless of which of the above two goals one wants
1600  * to achieve, service trees need first to be updated to know whether
1601  * the in-service queue must be preempted. To have service trees
1602  * correctly updated, the in-service queue must be expired and
1603  * rescheduled, and bfqq must be scheduled too. This is one of the
1604  * most costly operations (in future versions, the scheduling
1605  * mechanism may be re-designed in such a way to make it possible to
1606  * know whether preemption is needed without needing to update service
1607  * trees). In addition, queue preemptions almost always cause random
1608  * I/O, which may in turn cause loss of throughput. Finally, there may
1609  * even be no in-service queue when the next function is invoked (so,
1610  * no queue to compare timestamps with). Because of these facts, the
1611  * next function adopts the following simple scheme to avoid costly
1612  * operations, too frequent preemptions and too many dependencies on
1613  * the state of the scheduler: it requests the expiration of the
1614  * in-service queue (unconditionally) only for queues that need to
1615  * recover a hole. Then it delegates to other parts of the code the
1616  * responsibility of handling the above case 2.
1617  */
bfq_bfqq_update_budg_for_activation(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool arrived_in_time)1618 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1619 						struct bfq_queue *bfqq,
1620 						bool arrived_in_time)
1621 {
1622 	struct bfq_entity *entity = &bfqq->entity;
1623 
1624 	/*
1625 	 * In the next compound condition, we check also whether there
1626 	 * is some budget left, because otherwise there is no point in
1627 	 * trying to go on serving bfqq with this same budget: bfqq
1628 	 * would be expired immediately after being selected for
1629 	 * service. This would only cause useless overhead.
1630 	 */
1631 	if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1632 	    bfq_bfqq_budget_left(bfqq) > 0) {
1633 		/*
1634 		 * We do not clear the flag non_blocking_wait_rq here, as
1635 		 * the latter is used in bfq_activate_bfqq to signal
1636 		 * that timestamps need to be back-shifted (and is
1637 		 * cleared right after).
1638 		 */
1639 
1640 		/*
1641 		 * In next assignment we rely on that either
1642 		 * entity->service or entity->budget are not updated
1643 		 * on expiration if bfqq is empty (see
1644 		 * __bfq_bfqq_recalc_budget). Thus both quantities
1645 		 * remain unchanged after such an expiration, and the
1646 		 * following statement therefore assigns to
1647 		 * entity->budget the remaining budget on such an
1648 		 * expiration.
1649 		 */
1650 		entity->budget = min_t(unsigned long,
1651 				       bfq_bfqq_budget_left(bfqq),
1652 				       bfqq->max_budget);
1653 
1654 		/*
1655 		 * At this point, we have used entity->service to get
1656 		 * the budget left (needed for updating
1657 		 * entity->budget). Thus we finally can, and have to,
1658 		 * reset entity->service. The latter must be reset
1659 		 * because bfqq would otherwise be charged again for
1660 		 * the service it has received during its previous
1661 		 * service slot(s).
1662 		 */
1663 		entity->service = 0;
1664 
1665 		return true;
1666 	}
1667 
1668 	/*
1669 	 * We can finally complete expiration, by setting service to 0.
1670 	 */
1671 	entity->service = 0;
1672 	entity->budget = max_t(unsigned long, bfqq->max_budget,
1673 			       bfq_serv_to_charge(bfqq->next_rq, bfqq));
1674 	bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1675 	return false;
1676 }
1677 
1678 /*
1679  * Return the farthest past time instant according to jiffies
1680  * macros.
1681  */
bfq_smallest_from_now(void)1682 static unsigned long bfq_smallest_from_now(void)
1683 {
1684 	return jiffies - MAX_JIFFY_OFFSET;
1685 }
1686 
bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data * bfqd,struct bfq_queue * bfqq,unsigned int old_wr_coeff,bool wr_or_deserves_wr,bool interactive,bool in_burst,bool soft_rt)1687 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1688 					     struct bfq_queue *bfqq,
1689 					     unsigned int old_wr_coeff,
1690 					     bool wr_or_deserves_wr,
1691 					     bool interactive,
1692 					     bool in_burst,
1693 					     bool soft_rt)
1694 {
1695 	if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1696 		/* start a weight-raising period */
1697 		if (interactive) {
1698 			bfqq->service_from_wr = 0;
1699 			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1700 			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1701 		} else {
1702 			/*
1703 			 * No interactive weight raising in progress
1704 			 * here: assign minus infinity to
1705 			 * wr_start_at_switch_to_srt, to make sure
1706 			 * that, at the end of the soft-real-time
1707 			 * weight raising periods that is starting
1708 			 * now, no interactive weight-raising period
1709 			 * may be wrongly considered as still in
1710 			 * progress (and thus actually started by
1711 			 * mistake).
1712 			 */
1713 			bfqq->wr_start_at_switch_to_srt =
1714 				bfq_smallest_from_now();
1715 			bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1716 				BFQ_SOFTRT_WEIGHT_FACTOR;
1717 			bfqq->wr_cur_max_time =
1718 				bfqd->bfq_wr_rt_max_time;
1719 		}
1720 
1721 		/*
1722 		 * If needed, further reduce budget to make sure it is
1723 		 * close to bfqq's backlog, so as to reduce the
1724 		 * scheduling-error component due to a too large
1725 		 * budget. Do not care about throughput consequences,
1726 		 * but only about latency. Finally, do not assign a
1727 		 * too small budget either, to avoid increasing
1728 		 * latency by causing too frequent expirations.
1729 		 */
1730 		bfqq->entity.budget = min_t(unsigned long,
1731 					    bfqq->entity.budget,
1732 					    2 * bfq_min_budget(bfqd));
1733 	} else if (old_wr_coeff > 1) {
1734 		if (interactive) { /* update wr coeff and duration */
1735 			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1736 			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1737 		} else if (in_burst)
1738 			bfqq->wr_coeff = 1;
1739 		else if (soft_rt) {
1740 			/*
1741 			 * The application is now or still meeting the
1742 			 * requirements for being deemed soft rt.  We
1743 			 * can then correctly and safely (re)charge
1744 			 * the weight-raising duration for the
1745 			 * application with the weight-raising
1746 			 * duration for soft rt applications.
1747 			 *
1748 			 * In particular, doing this recharge now, i.e.,
1749 			 * before the weight-raising period for the
1750 			 * application finishes, reduces the probability
1751 			 * of the following negative scenario:
1752 			 * 1) the weight of a soft rt application is
1753 			 *    raised at startup (as for any newly
1754 			 *    created application),
1755 			 * 2) since the application is not interactive,
1756 			 *    at a certain time weight-raising is
1757 			 *    stopped for the application,
1758 			 * 3) at that time the application happens to
1759 			 *    still have pending requests, and hence
1760 			 *    is destined to not have a chance to be
1761 			 *    deemed soft rt before these requests are
1762 			 *    completed (see the comments to the
1763 			 *    function bfq_bfqq_softrt_next_start()
1764 			 *    for details on soft rt detection),
1765 			 * 4) these pending requests experience a high
1766 			 *    latency because the application is not
1767 			 *    weight-raised while they are pending.
1768 			 */
1769 			if (bfqq->wr_cur_max_time !=
1770 				bfqd->bfq_wr_rt_max_time) {
1771 				bfqq->wr_start_at_switch_to_srt =
1772 					bfqq->last_wr_start_finish;
1773 
1774 				bfqq->wr_cur_max_time =
1775 					bfqd->bfq_wr_rt_max_time;
1776 				bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1777 					BFQ_SOFTRT_WEIGHT_FACTOR;
1778 			}
1779 			bfqq->last_wr_start_finish = jiffies;
1780 		}
1781 	}
1782 }
1783 
bfq_bfqq_idle_for_long_time(struct bfq_data * bfqd,struct bfq_queue * bfqq)1784 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1785 					struct bfq_queue *bfqq)
1786 {
1787 	return bfqq->dispatched == 0 &&
1788 		time_is_before_jiffies(
1789 			bfqq->budget_timeout +
1790 			bfqd->bfq_wr_min_idle_time);
1791 }
1792 
1793 
1794 /*
1795  * Return true if bfqq is in a higher priority class, or has a higher
1796  * weight than the in-service queue.
1797  */
bfq_bfqq_higher_class_or_weight(struct bfq_queue * bfqq,struct bfq_queue * in_serv_bfqq)1798 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1799 					    struct bfq_queue *in_serv_bfqq)
1800 {
1801 	int bfqq_weight, in_serv_weight;
1802 
1803 	if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1804 		return true;
1805 
1806 	if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1807 		bfqq_weight = bfqq->entity.weight;
1808 		in_serv_weight = in_serv_bfqq->entity.weight;
1809 	} else {
1810 		if (bfqq->entity.parent)
1811 			bfqq_weight = bfqq->entity.parent->weight;
1812 		else
1813 			bfqq_weight = bfqq->entity.weight;
1814 		if (in_serv_bfqq->entity.parent)
1815 			in_serv_weight = in_serv_bfqq->entity.parent->weight;
1816 		else
1817 			in_serv_weight = in_serv_bfqq->entity.weight;
1818 	}
1819 
1820 	return bfqq_weight > in_serv_weight;
1821 }
1822 
1823 static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1824 
bfq_bfqq_handle_idle_busy_switch(struct bfq_data * bfqd,struct bfq_queue * bfqq,int old_wr_coeff,struct request * rq,bool * interactive)1825 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1826 					     struct bfq_queue *bfqq,
1827 					     int old_wr_coeff,
1828 					     struct request *rq,
1829 					     bool *interactive)
1830 {
1831 	bool soft_rt, in_burst,	wr_or_deserves_wr,
1832 		bfqq_wants_to_preempt,
1833 		idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1834 		/*
1835 		 * See the comments on
1836 		 * bfq_bfqq_update_budg_for_activation for
1837 		 * details on the usage of the next variable.
1838 		 */
1839 		arrived_in_time =  ktime_get_ns() <=
1840 			bfqq->ttime.last_end_request +
1841 			bfqd->bfq_slice_idle * 3;
1842 
1843 
1844 	/*
1845 	 * bfqq deserves to be weight-raised if:
1846 	 * - it is sync,
1847 	 * - it does not belong to a large burst,
1848 	 * - it has been idle for enough time or is soft real-time,
1849 	 * - is linked to a bfq_io_cq (it is not shared in any sense),
1850 	 * - has a default weight (otherwise we assume the user wanted
1851 	 *   to control its weight explicitly)
1852 	 */
1853 	in_burst = bfq_bfqq_in_large_burst(bfqq);
1854 	soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1855 		!BFQQ_TOTALLY_SEEKY(bfqq) &&
1856 		!in_burst &&
1857 		time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1858 		bfqq->dispatched == 0 &&
1859 		bfqq->entity.new_weight == 40;
1860 	*interactive = !in_burst && idle_for_long_time &&
1861 		bfqq->entity.new_weight == 40;
1862 	/*
1863 	 * Merged bfq_queues are kept out of weight-raising
1864 	 * (low-latency) mechanisms. The reason is that these queues
1865 	 * are usually created for non-interactive and
1866 	 * non-soft-real-time tasks. Yet this is not the case for
1867 	 * stably-merged queues. These queues are merged just because
1868 	 * they are created shortly after each other. So they may
1869 	 * easily serve the I/O of an interactive or soft-real time
1870 	 * application, if the application happens to spawn multiple
1871 	 * processes. So let also stably-merged queued enjoy weight
1872 	 * raising.
1873 	 */
1874 	wr_or_deserves_wr = bfqd->low_latency &&
1875 		(bfqq->wr_coeff > 1 ||
1876 		 (bfq_bfqq_sync(bfqq) &&
1877 		  (bfqq->bic || RQ_BIC(rq)->stably_merged) &&
1878 		   (*interactive || soft_rt)));
1879 
1880 	/*
1881 	 * Using the last flag, update budget and check whether bfqq
1882 	 * may want to preempt the in-service queue.
1883 	 */
1884 	bfqq_wants_to_preempt =
1885 		bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1886 						    arrived_in_time);
1887 
1888 	/*
1889 	 * If bfqq happened to be activated in a burst, but has been
1890 	 * idle for much more than an interactive queue, then we
1891 	 * assume that, in the overall I/O initiated in the burst, the
1892 	 * I/O associated with bfqq is finished. So bfqq does not need
1893 	 * to be treated as a queue belonging to a burst
1894 	 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1895 	 * if set, and remove bfqq from the burst list if it's
1896 	 * there. We do not decrement burst_size, because the fact
1897 	 * that bfqq does not need to belong to the burst list any
1898 	 * more does not invalidate the fact that bfqq was created in
1899 	 * a burst.
1900 	 */
1901 	if (likely(!bfq_bfqq_just_created(bfqq)) &&
1902 	    idle_for_long_time &&
1903 	    time_is_before_jiffies(
1904 		    bfqq->budget_timeout +
1905 		    msecs_to_jiffies(10000))) {
1906 		hlist_del_init(&bfqq->burst_list_node);
1907 		bfq_clear_bfqq_in_large_burst(bfqq);
1908 	}
1909 
1910 	bfq_clear_bfqq_just_created(bfqq);
1911 
1912 	if (bfqd->low_latency) {
1913 		if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1914 			/* wraparound */
1915 			bfqq->split_time =
1916 				jiffies - bfqd->bfq_wr_min_idle_time - 1;
1917 
1918 		if (time_is_before_jiffies(bfqq->split_time +
1919 					   bfqd->bfq_wr_min_idle_time)) {
1920 			bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1921 							 old_wr_coeff,
1922 							 wr_or_deserves_wr,
1923 							 *interactive,
1924 							 in_burst,
1925 							 soft_rt);
1926 
1927 			if (old_wr_coeff != bfqq->wr_coeff)
1928 				bfqq->entity.prio_changed = 1;
1929 		}
1930 	}
1931 
1932 	bfqq->last_idle_bklogged = jiffies;
1933 	bfqq->service_from_backlogged = 0;
1934 	bfq_clear_bfqq_softrt_update(bfqq);
1935 
1936 	bfq_add_bfqq_busy(bfqq);
1937 
1938 	/*
1939 	 * Expire in-service queue if preemption may be needed for
1940 	 * guarantees or throughput. As for guarantees, we care
1941 	 * explicitly about two cases. The first is that bfqq has to
1942 	 * recover a service hole, as explained in the comments on
1943 	 * bfq_bfqq_update_budg_for_activation(), i.e., that
1944 	 * bfqq_wants_to_preempt is true. However, if bfqq does not
1945 	 * carry time-critical I/O, then bfqq's bandwidth is less
1946 	 * important than that of queues that carry time-critical I/O.
1947 	 * So, as a further constraint, we consider this case only if
1948 	 * bfqq is at least as weight-raised, i.e., at least as time
1949 	 * critical, as the in-service queue.
1950 	 *
1951 	 * The second case is that bfqq is in a higher priority class,
1952 	 * or has a higher weight than the in-service queue. If this
1953 	 * condition does not hold, we don't care because, even if
1954 	 * bfqq does not start to be served immediately, the resulting
1955 	 * delay for bfqq's I/O is however lower or much lower than
1956 	 * the ideal completion time to be guaranteed to bfqq's I/O.
1957 	 *
1958 	 * In both cases, preemption is needed only if, according to
1959 	 * the timestamps of both bfqq and of the in-service queue,
1960 	 * bfqq actually is the next queue to serve. So, to reduce
1961 	 * useless preemptions, the return value of
1962 	 * next_queue_may_preempt() is considered in the next compound
1963 	 * condition too. Yet next_queue_may_preempt() just checks a
1964 	 * simple, necessary condition for bfqq to be the next queue
1965 	 * to serve. In fact, to evaluate a sufficient condition, the
1966 	 * timestamps of the in-service queue would need to be
1967 	 * updated, and this operation is quite costly (see the
1968 	 * comments on bfq_bfqq_update_budg_for_activation()).
1969 	 *
1970 	 * As for throughput, we ask bfq_better_to_idle() whether we
1971 	 * still need to plug I/O dispatching. If bfq_better_to_idle()
1972 	 * says no, then plugging is not needed any longer, either to
1973 	 * boost throughput or to perserve service guarantees. Then
1974 	 * the best option is to stop plugging I/O, as not doing so
1975 	 * would certainly lower throughput. We may end up in this
1976 	 * case if: (1) upon a dispatch attempt, we detected that it
1977 	 * was better to plug I/O dispatch, and to wait for a new
1978 	 * request to arrive for the currently in-service queue, but
1979 	 * (2) this switch of bfqq to busy changes the scenario.
1980 	 */
1981 	if (bfqd->in_service_queue &&
1982 	    ((bfqq_wants_to_preempt &&
1983 	      bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1984 	     bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1985 	     !bfq_better_to_idle(bfqd->in_service_queue)) &&
1986 	    next_queue_may_preempt(bfqd))
1987 		bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1988 				false, BFQQE_PREEMPTED);
1989 }
1990 
bfq_reset_inject_limit(struct bfq_data * bfqd,struct bfq_queue * bfqq)1991 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1992 				   struct bfq_queue *bfqq)
1993 {
1994 	/* invalidate baseline total service time */
1995 	bfqq->last_serv_time_ns = 0;
1996 
1997 	/*
1998 	 * Reset pointer in case we are waiting for
1999 	 * some request completion.
2000 	 */
2001 	bfqd->waited_rq = NULL;
2002 
2003 	/*
2004 	 * If bfqq has a short think time, then start by setting the
2005 	 * inject limit to 0 prudentially, because the service time of
2006 	 * an injected I/O request may be higher than the think time
2007 	 * of bfqq, and therefore, if one request was injected when
2008 	 * bfqq remains empty, this injected request might delay the
2009 	 * service of the next I/O request for bfqq significantly. In
2010 	 * case bfqq can actually tolerate some injection, then the
2011 	 * adaptive update will however raise the limit soon. This
2012 	 * lucky circumstance holds exactly because bfqq has a short
2013 	 * think time, and thus, after remaining empty, is likely to
2014 	 * get new I/O enqueued---and then completed---before being
2015 	 * expired. This is the very pattern that gives the
2016 	 * limit-update algorithm the chance to measure the effect of
2017 	 * injection on request service times, and then to update the
2018 	 * limit accordingly.
2019 	 *
2020 	 * However, in the following special case, the inject limit is
2021 	 * left to 1 even if the think time is short: bfqq's I/O is
2022 	 * synchronized with that of some other queue, i.e., bfqq may
2023 	 * receive new I/O only after the I/O of the other queue is
2024 	 * completed. Keeping the inject limit to 1 allows the
2025 	 * blocking I/O to be served while bfqq is in service. And
2026 	 * this is very convenient both for bfqq and for overall
2027 	 * throughput, as explained in detail in the comments in
2028 	 * bfq_update_has_short_ttime().
2029 	 *
2030 	 * On the opposite end, if bfqq has a long think time, then
2031 	 * start directly by 1, because:
2032 	 * a) on the bright side, keeping at most one request in
2033 	 * service in the drive is unlikely to cause any harm to the
2034 	 * latency of bfqq's requests, as the service time of a single
2035 	 * request is likely to be lower than the think time of bfqq;
2036 	 * b) on the downside, after becoming empty, bfqq is likely to
2037 	 * expire before getting its next request. With this request
2038 	 * arrival pattern, it is very hard to sample total service
2039 	 * times and update the inject limit accordingly (see comments
2040 	 * on bfq_update_inject_limit()). So the limit is likely to be
2041 	 * never, or at least seldom, updated.  As a consequence, by
2042 	 * setting the limit to 1, we avoid that no injection ever
2043 	 * occurs with bfqq. On the downside, this proactive step
2044 	 * further reduces chances to actually compute the baseline
2045 	 * total service time. Thus it reduces chances to execute the
2046 	 * limit-update algorithm and possibly raise the limit to more
2047 	 * than 1.
2048 	 */
2049 	if (bfq_bfqq_has_short_ttime(bfqq))
2050 		bfqq->inject_limit = 0;
2051 	else
2052 		bfqq->inject_limit = 1;
2053 
2054 	bfqq->decrease_time_jif = jiffies;
2055 }
2056 
bfq_update_io_intensity(struct bfq_queue * bfqq,u64 now_ns)2057 static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
2058 {
2059 	u64 tot_io_time = now_ns - bfqq->io_start_time;
2060 
2061 	if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2062 		bfqq->tot_idle_time +=
2063 			now_ns - bfqq->ttime.last_end_request;
2064 
2065 	if (unlikely(bfq_bfqq_just_created(bfqq)))
2066 		return;
2067 
2068 	/*
2069 	 * Must be busy for at least about 80% of the time to be
2070 	 * considered I/O bound.
2071 	 */
2072 	if (bfqq->tot_idle_time * 5 > tot_io_time)
2073 		bfq_clear_bfqq_IO_bound(bfqq);
2074 	else
2075 		bfq_mark_bfqq_IO_bound(bfqq);
2076 
2077 	/*
2078 	 * Keep an observation window of at most 200 ms in the past
2079 	 * from now.
2080 	 */
2081 	if (tot_io_time > 200 * NSEC_PER_MSEC) {
2082 		bfqq->io_start_time = now_ns - (tot_io_time>>1);
2083 		bfqq->tot_idle_time >>= 1;
2084 	}
2085 }
2086 
2087 /*
2088  * Detect whether bfqq's I/O seems synchronized with that of some
2089  * other queue, i.e., whether bfqq, after remaining empty, happens to
2090  * receive new I/O only right after some I/O request of the other
2091  * queue has been completed. We call waker queue the other queue, and
2092  * we assume, for simplicity, that bfqq may have at most one waker
2093  * queue.
2094  *
2095  * A remarkable throughput boost can be reached by unconditionally
2096  * injecting the I/O of the waker queue, every time a new
2097  * bfq_dispatch_request happens to be invoked while I/O is being
2098  * plugged for bfqq.  In addition to boosting throughput, this
2099  * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2100  * bfqq. Note that these same results may be achieved with the general
2101  * injection mechanism, but less effectively. For details on this
2102  * aspect, see the comments on the choice of the queue for injection
2103  * in bfq_select_queue().
2104  *
2105  * Turning back to the detection of a waker queue, a queue Q is deemed as a
2106  * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2107  * non empty right after a request of Q has been completed within given
2108  * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2109  * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2110  * still being served by the drive, and may receive new I/O on the completion
2111  * of some of the in-flight requests. In particular, on the first time, Q is
2112  * tentatively set as a candidate waker queue, while on the third consecutive
2113  * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2114  * is a waker queue for bfqq. These detection steps are performed only if bfqq
2115  * has a long think time, so as to make it more likely that bfqq's I/O is
2116  * actually being blocked by a synchronization. This last filter, plus the
2117  * above three-times requirement and time limit for detection, make false
2118  * positives less likely.
2119  *
2120  * NOTE
2121  *
2122  * The sooner a waker queue is detected, the sooner throughput can be
2123  * boosted by injecting I/O from the waker queue. Fortunately,
2124  * detection is likely to be actually fast, for the following
2125  * reasons. While blocked by synchronization, bfqq has a long think
2126  * time. This implies that bfqq's inject limit is at least equal to 1
2127  * (see the comments in bfq_update_inject_limit()). So, thanks to
2128  * injection, the waker queue is likely to be served during the very
2129  * first I/O-plugging time interval for bfqq. This triggers the first
2130  * step of the detection mechanism. Thanks again to injection, the
2131  * candidate waker queue is then likely to be confirmed no later than
2132  * during the next I/O-plugging interval for bfqq.
2133  *
2134  * ISSUE
2135  *
2136  * On queue merging all waker information is lost.
2137  */
bfq_check_waker(struct bfq_data * bfqd,struct bfq_queue * bfqq,u64 now_ns)2138 static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2139 			    u64 now_ns)
2140 {
2141 	char waker_name[MAX_BFQQ_NAME_LENGTH];
2142 
2143 	if (!bfqd->last_completed_rq_bfqq ||
2144 	    bfqd->last_completed_rq_bfqq == bfqq ||
2145 	    bfq_bfqq_has_short_ttime(bfqq) ||
2146 	    now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC)
2147 		return;
2148 
2149 	/*
2150 	 * We reset waker detection logic also if too much time has passed
2151  	 * since the first detection. If wakeups are rare, pointless idling
2152 	 * doesn't hurt throughput that much. The condition below makes sure
2153 	 * we do not uselessly idle blocking waker in more than 1/64 cases.
2154 	 */
2155 	if (bfqd->last_completed_rq_bfqq !=
2156 	    bfqq->tentative_waker_bfqq ||
2157 	    now_ns > bfqq->waker_detection_started +
2158 					128 * (u64)bfqd->bfq_slice_idle) {
2159 		/*
2160 		 * First synchronization detected with a
2161 		 * candidate waker queue, or with a different
2162 		 * candidate waker queue from the current one.
2163 		 */
2164 		bfqq->tentative_waker_bfqq =
2165 			bfqd->last_completed_rq_bfqq;
2166 		bfqq->num_waker_detections = 1;
2167 		bfqq->waker_detection_started = now_ns;
2168 		bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2169 			      MAX_BFQQ_NAME_LENGTH);
2170 		bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2171 	} else /* Same tentative waker queue detected again */
2172 		bfqq->num_waker_detections++;
2173 
2174 	if (bfqq->num_waker_detections == 3) {
2175 		bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2176 		bfqq->tentative_waker_bfqq = NULL;
2177 		bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2178 			      MAX_BFQQ_NAME_LENGTH);
2179 		bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2180 
2181 		/*
2182 		 * If the waker queue disappears, then
2183 		 * bfqq->waker_bfqq must be reset. To
2184 		 * this goal, we maintain in each
2185 		 * waker queue a list, woken_list, of
2186 		 * all the queues that reference the
2187 		 * waker queue through their
2188 		 * waker_bfqq pointer. When the waker
2189 		 * queue exits, the waker_bfqq pointer
2190 		 * of all the queues in the woken_list
2191 		 * is reset.
2192 		 *
2193 		 * In addition, if bfqq is already in
2194 		 * the woken_list of a waker queue,
2195 		 * then, before being inserted into
2196 		 * the woken_list of a new waker
2197 		 * queue, bfqq must be removed from
2198 		 * the woken_list of the old waker
2199 		 * queue.
2200 		 */
2201 		if (!hlist_unhashed(&bfqq->woken_list_node))
2202 			hlist_del_init(&bfqq->woken_list_node);
2203 		hlist_add_head(&bfqq->woken_list_node,
2204 			       &bfqd->last_completed_rq_bfqq->woken_list);
2205 	}
2206 }
2207 
bfq_add_request(struct request * rq)2208 static void bfq_add_request(struct request *rq)
2209 {
2210 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
2211 	struct bfq_data *bfqd = bfqq->bfqd;
2212 	struct request *next_rq, *prev;
2213 	unsigned int old_wr_coeff = bfqq->wr_coeff;
2214 	bool interactive = false;
2215 	u64 now_ns = ktime_get_ns();
2216 
2217 	bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2218 	bfqq->queued[rq_is_sync(rq)]++;
2219 	/*
2220 	 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2221 	 * may be read without holding the lock in bfq_has_work().
2222 	 */
2223 	WRITE_ONCE(bfqd->queued, bfqd->queued + 1);
2224 
2225 	if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) {
2226 		bfq_check_waker(bfqd, bfqq, now_ns);
2227 
2228 		/*
2229 		 * Periodically reset inject limit, to make sure that
2230 		 * the latter eventually drops in case workload
2231 		 * changes, see step (3) in the comments on
2232 		 * bfq_update_inject_limit().
2233 		 */
2234 		if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2235 					     msecs_to_jiffies(1000)))
2236 			bfq_reset_inject_limit(bfqd, bfqq);
2237 
2238 		/*
2239 		 * The following conditions must hold to setup a new
2240 		 * sampling of total service time, and then a new
2241 		 * update of the inject limit:
2242 		 * - bfqq is in service, because the total service
2243 		 *   time is evaluated only for the I/O requests of
2244 		 *   the queues in service;
2245 		 * - this is the right occasion to compute or to
2246 		 *   lower the baseline total service time, because
2247 		 *   there are actually no requests in the drive,
2248 		 *   or
2249 		 *   the baseline total service time is available, and
2250 		 *   this is the right occasion to compute the other
2251 		 *   quantity needed to update the inject limit, i.e.,
2252 		 *   the total service time caused by the amount of
2253 		 *   injection allowed by the current value of the
2254 		 *   limit. It is the right occasion because injection
2255 		 *   has actually been performed during the service
2256 		 *   hole, and there are still in-flight requests,
2257 		 *   which are very likely to be exactly the injected
2258 		 *   requests, or part of them;
2259 		 * - the minimum interval for sampling the total
2260 		 *   service time and updating the inject limit has
2261 		 *   elapsed.
2262 		 */
2263 		if (bfqq == bfqd->in_service_queue &&
2264 		    (bfqd->rq_in_driver == 0 ||
2265 		     (bfqq->last_serv_time_ns > 0 &&
2266 		      bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2267 		    time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2268 					      msecs_to_jiffies(10))) {
2269 			bfqd->last_empty_occupied_ns = ktime_get_ns();
2270 			/*
2271 			 * Start the state machine for measuring the
2272 			 * total service time of rq: setting
2273 			 * wait_dispatch will cause bfqd->waited_rq to
2274 			 * be set when rq will be dispatched.
2275 			 */
2276 			bfqd->wait_dispatch = true;
2277 			/*
2278 			 * If there is no I/O in service in the drive,
2279 			 * then possible injection occurred before the
2280 			 * arrival of rq will not affect the total
2281 			 * service time of rq. So the injection limit
2282 			 * must not be updated as a function of such
2283 			 * total service time, unless new injection
2284 			 * occurs before rq is completed. To have the
2285 			 * injection limit updated only in the latter
2286 			 * case, reset rqs_injected here (rqs_injected
2287 			 * will be set in case injection is performed
2288 			 * on bfqq before rq is completed).
2289 			 */
2290 			if (bfqd->rq_in_driver == 0)
2291 				bfqd->rqs_injected = false;
2292 		}
2293 	}
2294 
2295 	if (bfq_bfqq_sync(bfqq))
2296 		bfq_update_io_intensity(bfqq, now_ns);
2297 
2298 	elv_rb_add(&bfqq->sort_list, rq);
2299 
2300 	/*
2301 	 * Check if this request is a better next-serve candidate.
2302 	 */
2303 	prev = bfqq->next_rq;
2304 	next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2305 	bfqq->next_rq = next_rq;
2306 
2307 	/*
2308 	 * Adjust priority tree position, if next_rq changes.
2309 	 * See comments on bfq_pos_tree_add_move() for the unlikely().
2310 	 */
2311 	if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2312 		bfq_pos_tree_add_move(bfqd, bfqq);
2313 
2314 	if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2315 		bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2316 						 rq, &interactive);
2317 	else {
2318 		if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2319 		    time_is_before_jiffies(
2320 				bfqq->last_wr_start_finish +
2321 				bfqd->bfq_wr_min_inter_arr_async)) {
2322 			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2323 			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2324 
2325 			bfqd->wr_busy_queues++;
2326 			bfqq->entity.prio_changed = 1;
2327 		}
2328 		if (prev != bfqq->next_rq)
2329 			bfq_updated_next_req(bfqd, bfqq);
2330 	}
2331 
2332 	/*
2333 	 * Assign jiffies to last_wr_start_finish in the following
2334 	 * cases:
2335 	 *
2336 	 * . if bfqq is not going to be weight-raised, because, for
2337 	 *   non weight-raised queues, last_wr_start_finish stores the
2338 	 *   arrival time of the last request; as of now, this piece
2339 	 *   of information is used only for deciding whether to
2340 	 *   weight-raise async queues
2341 	 *
2342 	 * . if bfqq is not weight-raised, because, if bfqq is now
2343 	 *   switching to weight-raised, then last_wr_start_finish
2344 	 *   stores the time when weight-raising starts
2345 	 *
2346 	 * . if bfqq is interactive, because, regardless of whether
2347 	 *   bfqq is currently weight-raised, the weight-raising
2348 	 *   period must start or restart (this case is considered
2349 	 *   separately because it is not detected by the above
2350 	 *   conditions, if bfqq is already weight-raised)
2351 	 *
2352 	 * last_wr_start_finish has to be updated also if bfqq is soft
2353 	 * real-time, because the weight-raising period is constantly
2354 	 * restarted on idle-to-busy transitions for these queues, but
2355 	 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2356 	 * needed.
2357 	 */
2358 	if (bfqd->low_latency &&
2359 		(old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2360 		bfqq->last_wr_start_finish = jiffies;
2361 }
2362 
bfq_find_rq_fmerge(struct bfq_data * bfqd,struct bio * bio,struct request_queue * q)2363 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2364 					  struct bio *bio,
2365 					  struct request_queue *q)
2366 {
2367 	struct bfq_queue *bfqq = bfqd->bio_bfqq;
2368 
2369 
2370 	if (bfqq)
2371 		return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2372 
2373 	return NULL;
2374 }
2375 
get_sdist(sector_t last_pos,struct request * rq)2376 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2377 {
2378 	if (last_pos)
2379 		return abs(blk_rq_pos(rq) - last_pos);
2380 
2381 	return 0;
2382 }
2383 
2384 #if 0 /* Still not clear if we can do without next two functions */
2385 static void bfq_activate_request(struct request_queue *q, struct request *rq)
2386 {
2387 	struct bfq_data *bfqd = q->elevator->elevator_data;
2388 
2389 	bfqd->rq_in_driver++;
2390 }
2391 
2392 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
2393 {
2394 	struct bfq_data *bfqd = q->elevator->elevator_data;
2395 
2396 	bfqd->rq_in_driver--;
2397 }
2398 #endif
2399 
bfq_remove_request(struct request_queue * q,struct request * rq)2400 static void bfq_remove_request(struct request_queue *q,
2401 			       struct request *rq)
2402 {
2403 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
2404 	struct bfq_data *bfqd = bfqq->bfqd;
2405 	const int sync = rq_is_sync(rq);
2406 
2407 	if (bfqq->next_rq == rq) {
2408 		bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2409 		bfq_updated_next_req(bfqd, bfqq);
2410 	}
2411 
2412 	if (rq->queuelist.prev != &rq->queuelist)
2413 		list_del_init(&rq->queuelist);
2414 	bfqq->queued[sync]--;
2415 	/*
2416 	 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2417 	 * may be read without holding the lock in bfq_has_work().
2418 	 */
2419 	WRITE_ONCE(bfqd->queued, bfqd->queued - 1);
2420 	elv_rb_del(&bfqq->sort_list, rq);
2421 
2422 	elv_rqhash_del(q, rq);
2423 	if (q->last_merge == rq)
2424 		q->last_merge = NULL;
2425 
2426 	if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2427 		bfqq->next_rq = NULL;
2428 
2429 		if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2430 			bfq_del_bfqq_busy(bfqq, false);
2431 			/*
2432 			 * bfqq emptied. In normal operation, when
2433 			 * bfqq is empty, bfqq->entity.service and
2434 			 * bfqq->entity.budget must contain,
2435 			 * respectively, the service received and the
2436 			 * budget used last time bfqq emptied. These
2437 			 * facts do not hold in this case, as at least
2438 			 * this last removal occurred while bfqq is
2439 			 * not in service. To avoid inconsistencies,
2440 			 * reset both bfqq->entity.service and
2441 			 * bfqq->entity.budget, if bfqq has still a
2442 			 * process that may issue I/O requests to it.
2443 			 */
2444 			bfqq->entity.budget = bfqq->entity.service = 0;
2445 		}
2446 
2447 		/*
2448 		 * Remove queue from request-position tree as it is empty.
2449 		 */
2450 		if (bfqq->pos_root) {
2451 			rb_erase(&bfqq->pos_node, bfqq->pos_root);
2452 			bfqq->pos_root = NULL;
2453 		}
2454 	} else {
2455 		/* see comments on bfq_pos_tree_add_move() for the unlikely() */
2456 		if (unlikely(!bfqd->nonrot_with_queueing))
2457 			bfq_pos_tree_add_move(bfqd, bfqq);
2458 	}
2459 
2460 	if (rq->cmd_flags & REQ_META)
2461 		bfqq->meta_pending--;
2462 
2463 }
2464 
bfq_bio_merge(struct request_queue * q,struct bio * bio,unsigned int nr_segs)2465 static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2466 		unsigned int nr_segs)
2467 {
2468 	struct bfq_data *bfqd = q->elevator->elevator_data;
2469 	struct request *free = NULL;
2470 	/*
2471 	 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2472 	 * store its return value for later use, to avoid nesting
2473 	 * queue_lock inside the bfqd->lock. We assume that the bic
2474 	 * returned by bfq_bic_lookup does not go away before
2475 	 * bfqd->lock is taken.
2476 	 */
2477 	struct bfq_io_cq *bic = bfq_bic_lookup(q);
2478 	bool ret;
2479 
2480 	spin_lock_irq(&bfqd->lock);
2481 
2482 	if (bic) {
2483 		/*
2484 		 * Make sure cgroup info is uptodate for current process before
2485 		 * considering the merge.
2486 		 */
2487 		bfq_bic_update_cgroup(bic, bio);
2488 
2489 		bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2490 	} else {
2491 		bfqd->bio_bfqq = NULL;
2492 	}
2493 	bfqd->bio_bic = bic;
2494 
2495 	ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2496 
2497 	spin_unlock_irq(&bfqd->lock);
2498 	if (free)
2499 		blk_mq_free_request(free);
2500 
2501 	return ret;
2502 }
2503 
bfq_request_merge(struct request_queue * q,struct request ** req,struct bio * bio)2504 static int bfq_request_merge(struct request_queue *q, struct request **req,
2505 			     struct bio *bio)
2506 {
2507 	struct bfq_data *bfqd = q->elevator->elevator_data;
2508 	struct request *__rq;
2509 
2510 	__rq = bfq_find_rq_fmerge(bfqd, bio, q);
2511 	if (__rq && elv_bio_merge_ok(__rq, bio)) {
2512 		*req = __rq;
2513 
2514 		if (blk_discard_mergable(__rq))
2515 			return ELEVATOR_DISCARD_MERGE;
2516 		return ELEVATOR_FRONT_MERGE;
2517 	}
2518 
2519 	return ELEVATOR_NO_MERGE;
2520 }
2521 
bfq_request_merged(struct request_queue * q,struct request * req,enum elv_merge type)2522 static void bfq_request_merged(struct request_queue *q, struct request *req,
2523 			       enum elv_merge type)
2524 {
2525 	if (type == ELEVATOR_FRONT_MERGE &&
2526 	    rb_prev(&req->rb_node) &&
2527 	    blk_rq_pos(req) <
2528 	    blk_rq_pos(container_of(rb_prev(&req->rb_node),
2529 				    struct request, rb_node))) {
2530 		struct bfq_queue *bfqq = RQ_BFQQ(req);
2531 		struct bfq_data *bfqd;
2532 		struct request *prev, *next_rq;
2533 
2534 		if (!bfqq)
2535 			return;
2536 
2537 		bfqd = bfqq->bfqd;
2538 
2539 		/* Reposition request in its sort_list */
2540 		elv_rb_del(&bfqq->sort_list, req);
2541 		elv_rb_add(&bfqq->sort_list, req);
2542 
2543 		/* Choose next request to be served for bfqq */
2544 		prev = bfqq->next_rq;
2545 		next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2546 					 bfqd->last_position);
2547 		bfqq->next_rq = next_rq;
2548 		/*
2549 		 * If next_rq changes, update both the queue's budget to
2550 		 * fit the new request and the queue's position in its
2551 		 * rq_pos_tree.
2552 		 */
2553 		if (prev != bfqq->next_rq) {
2554 			bfq_updated_next_req(bfqd, bfqq);
2555 			/*
2556 			 * See comments on bfq_pos_tree_add_move() for
2557 			 * the unlikely().
2558 			 */
2559 			if (unlikely(!bfqd->nonrot_with_queueing))
2560 				bfq_pos_tree_add_move(bfqd, bfqq);
2561 		}
2562 	}
2563 }
2564 
2565 /*
2566  * This function is called to notify the scheduler that the requests
2567  * rq and 'next' have been merged, with 'next' going away.  BFQ
2568  * exploits this hook to address the following issue: if 'next' has a
2569  * fifo_time lower that rq, then the fifo_time of rq must be set to
2570  * the value of 'next', to not forget the greater age of 'next'.
2571  *
2572  * NOTE: in this function we assume that rq is in a bfq_queue, basing
2573  * on that rq is picked from the hash table q->elevator->hash, which,
2574  * in its turn, is filled only with I/O requests present in
2575  * bfq_queues, while BFQ is in use for the request queue q. In fact,
2576  * the function that fills this hash table (elv_rqhash_add) is called
2577  * only by bfq_insert_request.
2578  */
bfq_requests_merged(struct request_queue * q,struct request * rq,struct request * next)2579 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2580 				struct request *next)
2581 {
2582 	struct bfq_queue *bfqq = RQ_BFQQ(rq),
2583 		*next_bfqq = RQ_BFQQ(next);
2584 
2585 	if (!bfqq)
2586 		goto remove;
2587 
2588 	/*
2589 	 * If next and rq belong to the same bfq_queue and next is older
2590 	 * than rq, then reposition rq in the fifo (by substituting next
2591 	 * with rq). Otherwise, if next and rq belong to different
2592 	 * bfq_queues, never reposition rq: in fact, we would have to
2593 	 * reposition it with respect to next's position in its own fifo,
2594 	 * which would most certainly be too expensive with respect to
2595 	 * the benefits.
2596 	 */
2597 	if (bfqq == next_bfqq &&
2598 	    !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2599 	    next->fifo_time < rq->fifo_time) {
2600 		list_del_init(&rq->queuelist);
2601 		list_replace_init(&next->queuelist, &rq->queuelist);
2602 		rq->fifo_time = next->fifo_time;
2603 	}
2604 
2605 	if (bfqq->next_rq == next)
2606 		bfqq->next_rq = rq;
2607 
2608 	bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2609 remove:
2610 	/* Merged request may be in the IO scheduler. Remove it. */
2611 	if (!RB_EMPTY_NODE(&next->rb_node)) {
2612 		bfq_remove_request(next->q, next);
2613 		if (next_bfqq)
2614 			bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2615 						    next->cmd_flags);
2616 	}
2617 }
2618 
2619 /* Must be called with bfqq != NULL */
bfq_bfqq_end_wr(struct bfq_queue * bfqq)2620 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2621 {
2622 	/*
2623 	 * If bfqq has been enjoying interactive weight-raising, then
2624 	 * reset soft_rt_next_start. We do it for the following
2625 	 * reason. bfqq may have been conveying the I/O needed to load
2626 	 * a soft real-time application. Such an application actually
2627 	 * exhibits a soft real-time I/O pattern after it finishes
2628 	 * loading, and finally starts doing its job. But, if bfqq has
2629 	 * been receiving a lot of bandwidth so far (likely to happen
2630 	 * on a fast device), then soft_rt_next_start now contains a
2631 	 * high value that. So, without this reset, bfqq would be
2632 	 * prevented from being possibly considered as soft_rt for a
2633 	 * very long time.
2634 	 */
2635 
2636 	if (bfqq->wr_cur_max_time !=
2637 	    bfqq->bfqd->bfq_wr_rt_max_time)
2638 		bfqq->soft_rt_next_start = jiffies;
2639 
2640 	if (bfq_bfqq_busy(bfqq))
2641 		bfqq->bfqd->wr_busy_queues--;
2642 	bfqq->wr_coeff = 1;
2643 	bfqq->wr_cur_max_time = 0;
2644 	bfqq->last_wr_start_finish = jiffies;
2645 	/*
2646 	 * Trigger a weight change on the next invocation of
2647 	 * __bfq_entity_update_weight_prio.
2648 	 */
2649 	bfqq->entity.prio_changed = 1;
2650 }
2651 
bfq_end_wr_async_queues(struct bfq_data * bfqd,struct bfq_group * bfqg)2652 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2653 			     struct bfq_group *bfqg)
2654 {
2655 	int i, j;
2656 
2657 	for (i = 0; i < 2; i++)
2658 		for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2659 			if (bfqg->async_bfqq[i][j])
2660 				bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2661 	if (bfqg->async_idle_bfqq)
2662 		bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2663 }
2664 
bfq_end_wr(struct bfq_data * bfqd)2665 static void bfq_end_wr(struct bfq_data *bfqd)
2666 {
2667 	struct bfq_queue *bfqq;
2668 
2669 	spin_lock_irq(&bfqd->lock);
2670 
2671 	list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2672 		bfq_bfqq_end_wr(bfqq);
2673 	list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2674 		bfq_bfqq_end_wr(bfqq);
2675 	bfq_end_wr_async(bfqd);
2676 
2677 	spin_unlock_irq(&bfqd->lock);
2678 }
2679 
bfq_io_struct_pos(void * io_struct,bool request)2680 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2681 {
2682 	if (request)
2683 		return blk_rq_pos(io_struct);
2684 	else
2685 		return ((struct bio *)io_struct)->bi_iter.bi_sector;
2686 }
2687 
bfq_rq_close_to_sector(void * io_struct,bool request,sector_t sector)2688 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2689 				  sector_t sector)
2690 {
2691 	return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2692 	       BFQQ_CLOSE_THR;
2693 }
2694 
bfqq_find_close(struct bfq_data * bfqd,struct bfq_queue * bfqq,sector_t sector)2695 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2696 					 struct bfq_queue *bfqq,
2697 					 sector_t sector)
2698 {
2699 	struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2700 	struct rb_node *parent, *node;
2701 	struct bfq_queue *__bfqq;
2702 
2703 	if (RB_EMPTY_ROOT(root))
2704 		return NULL;
2705 
2706 	/*
2707 	 * First, if we find a request starting at the end of the last
2708 	 * request, choose it.
2709 	 */
2710 	__bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2711 	if (__bfqq)
2712 		return __bfqq;
2713 
2714 	/*
2715 	 * If the exact sector wasn't found, the parent of the NULL leaf
2716 	 * will contain the closest sector (rq_pos_tree sorted by
2717 	 * next_request position).
2718 	 */
2719 	__bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2720 	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2721 		return __bfqq;
2722 
2723 	if (blk_rq_pos(__bfqq->next_rq) < sector)
2724 		node = rb_next(&__bfqq->pos_node);
2725 	else
2726 		node = rb_prev(&__bfqq->pos_node);
2727 	if (!node)
2728 		return NULL;
2729 
2730 	__bfqq = rb_entry(node, struct bfq_queue, pos_node);
2731 	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2732 		return __bfqq;
2733 
2734 	return NULL;
2735 }
2736 
bfq_find_close_cooperator(struct bfq_data * bfqd,struct bfq_queue * cur_bfqq,sector_t sector)2737 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2738 						   struct bfq_queue *cur_bfqq,
2739 						   sector_t sector)
2740 {
2741 	struct bfq_queue *bfqq;
2742 
2743 	/*
2744 	 * We shall notice if some of the queues are cooperating,
2745 	 * e.g., working closely on the same area of the device. In
2746 	 * that case, we can group them together and: 1) don't waste
2747 	 * time idling, and 2) serve the union of their requests in
2748 	 * the best possible order for throughput.
2749 	 */
2750 	bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2751 	if (!bfqq || bfqq == cur_bfqq)
2752 		return NULL;
2753 
2754 	return bfqq;
2755 }
2756 
2757 static struct bfq_queue *
bfq_setup_merge(struct bfq_queue * bfqq,struct bfq_queue * new_bfqq)2758 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2759 {
2760 	int process_refs, new_process_refs;
2761 	struct bfq_queue *__bfqq;
2762 
2763 	/*
2764 	 * If there are no process references on the new_bfqq, then it is
2765 	 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2766 	 * may have dropped their last reference (not just their last process
2767 	 * reference).
2768 	 */
2769 	if (!bfqq_process_refs(new_bfqq))
2770 		return NULL;
2771 
2772 	/* Avoid a circular list and skip interim queue merges. */
2773 	while ((__bfqq = new_bfqq->new_bfqq)) {
2774 		if (__bfqq == bfqq)
2775 			return NULL;
2776 		new_bfqq = __bfqq;
2777 	}
2778 
2779 	process_refs = bfqq_process_refs(bfqq);
2780 	new_process_refs = bfqq_process_refs(new_bfqq);
2781 	/*
2782 	 * If the process for the bfqq has gone away, there is no
2783 	 * sense in merging the queues.
2784 	 */
2785 	if (process_refs == 0 || new_process_refs == 0)
2786 		return NULL;
2787 
2788 	/*
2789 	 * Make sure merged queues belong to the same parent. Parents could
2790 	 * have changed since the time we decided the two queues are suitable
2791 	 * for merging.
2792 	 */
2793 	if (new_bfqq->entity.parent != bfqq->entity.parent)
2794 		return NULL;
2795 
2796 	bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2797 		new_bfqq->pid);
2798 
2799 	/*
2800 	 * Merging is just a redirection: the requests of the process
2801 	 * owning one of the two queues are redirected to the other queue.
2802 	 * The latter queue, in its turn, is set as shared if this is the
2803 	 * first time that the requests of some process are redirected to
2804 	 * it.
2805 	 *
2806 	 * We redirect bfqq to new_bfqq and not the opposite, because
2807 	 * we are in the context of the process owning bfqq, thus we
2808 	 * have the io_cq of this process. So we can immediately
2809 	 * configure this io_cq to redirect the requests of the
2810 	 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2811 	 * not available any more (new_bfqq->bic == NULL).
2812 	 *
2813 	 * Anyway, even in case new_bfqq coincides with the in-service
2814 	 * queue, redirecting requests the in-service queue is the
2815 	 * best option, as we feed the in-service queue with new
2816 	 * requests close to the last request served and, by doing so,
2817 	 * are likely to increase the throughput.
2818 	 */
2819 	bfqq->new_bfqq = new_bfqq;
2820 	/*
2821 	 * The above assignment schedules the following redirections:
2822 	 * each time some I/O for bfqq arrives, the process that
2823 	 * generated that I/O is disassociated from bfqq and
2824 	 * associated with new_bfqq. Here we increases new_bfqq->ref
2825 	 * in advance, adding the number of processes that are
2826 	 * expected to be associated with new_bfqq as they happen to
2827 	 * issue I/O.
2828 	 */
2829 	new_bfqq->ref += process_refs;
2830 	return new_bfqq;
2831 }
2832 
bfq_may_be_close_cooperator(struct bfq_queue * bfqq,struct bfq_queue * new_bfqq)2833 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2834 					struct bfq_queue *new_bfqq)
2835 {
2836 	if (bfq_too_late_for_merging(new_bfqq))
2837 		return false;
2838 
2839 	if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2840 	    (bfqq->ioprio_class != new_bfqq->ioprio_class))
2841 		return false;
2842 
2843 	/*
2844 	 * If either of the queues has already been detected as seeky,
2845 	 * then merging it with the other queue is unlikely to lead to
2846 	 * sequential I/O.
2847 	 */
2848 	if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2849 		return false;
2850 
2851 	/*
2852 	 * Interleaved I/O is known to be done by (some) applications
2853 	 * only for reads, so it does not make sense to merge async
2854 	 * queues.
2855 	 */
2856 	if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2857 		return false;
2858 
2859 	return true;
2860 }
2861 
2862 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2863 					     struct bfq_queue *bfqq);
2864 
2865 /*
2866  * Attempt to schedule a merge of bfqq with the currently in-service
2867  * queue or with a close queue among the scheduled queues.  Return
2868  * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2869  * structure otherwise.
2870  *
2871  * The OOM queue is not allowed to participate to cooperation: in fact, since
2872  * the requests temporarily redirected to the OOM queue could be redirected
2873  * again to dedicated queues at any time, the state needed to correctly
2874  * handle merging with the OOM queue would be quite complex and expensive
2875  * to maintain. Besides, in such a critical condition as an out of memory,
2876  * the benefits of queue merging may be little relevant, or even negligible.
2877  *
2878  * WARNING: queue merging may impair fairness among non-weight raised
2879  * queues, for at least two reasons: 1) the original weight of a
2880  * merged queue may change during the merged state, 2) even being the
2881  * weight the same, a merged queue may be bloated with many more
2882  * requests than the ones produced by its originally-associated
2883  * process.
2884  */
2885 static struct bfq_queue *
bfq_setup_cooperator(struct bfq_data * bfqd,struct bfq_queue * bfqq,void * io_struct,bool request,struct bfq_io_cq * bic)2886 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2887 		     void *io_struct, bool request, struct bfq_io_cq *bic)
2888 {
2889 	struct bfq_queue *in_service_bfqq, *new_bfqq;
2890 
2891 	/* if a merge has already been setup, then proceed with that first */
2892 	if (bfqq->new_bfqq)
2893 		return bfqq->new_bfqq;
2894 
2895 	/*
2896 	 * Check delayed stable merge for rotational or non-queueing
2897 	 * devs. For this branch to be executed, bfqq must not be
2898 	 * currently merged with some other queue (i.e., bfqq->bic
2899 	 * must be non null). If we considered also merged queues,
2900 	 * then we should also check whether bfqq has already been
2901 	 * merged with bic->stable_merge_bfqq. But this would be
2902 	 * costly and complicated.
2903 	 */
2904 	if (unlikely(!bfqd->nonrot_with_queueing)) {
2905 		/*
2906 		 * Make sure also that bfqq is sync, because
2907 		 * bic->stable_merge_bfqq may point to some queue (for
2908 		 * stable merging) also if bic is associated with a
2909 		 * sync queue, but this bfqq is async
2910 		 */
2911 		if (bfq_bfqq_sync(bfqq) && bic->stable_merge_bfqq &&
2912 		    !bfq_bfqq_just_created(bfqq) &&
2913 		    time_is_before_jiffies(bfqq->split_time +
2914 					  msecs_to_jiffies(bfq_late_stable_merging)) &&
2915 		    time_is_before_jiffies(bfqq->creation_time +
2916 					   msecs_to_jiffies(bfq_late_stable_merging))) {
2917 			struct bfq_queue *stable_merge_bfqq =
2918 				bic->stable_merge_bfqq;
2919 			int proc_ref = min(bfqq_process_refs(bfqq),
2920 					   bfqq_process_refs(stable_merge_bfqq));
2921 
2922 			/* deschedule stable merge, because done or aborted here */
2923 			bfq_put_stable_ref(stable_merge_bfqq);
2924 
2925 			bic->stable_merge_bfqq = NULL;
2926 
2927 			if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
2928 			    proc_ref > 0) {
2929 				/* next function will take at least one ref */
2930 				struct bfq_queue *new_bfqq =
2931 					bfq_setup_merge(bfqq, stable_merge_bfqq);
2932 
2933 				if (new_bfqq) {
2934 					bic->stably_merged = true;
2935 					if (new_bfqq->bic)
2936 						new_bfqq->bic->stably_merged =
2937 									true;
2938 				}
2939 				return new_bfqq;
2940 			} else
2941 				return NULL;
2942 		}
2943 	}
2944 
2945 	/*
2946 	 * Do not perform queue merging if the device is non
2947 	 * rotational and performs internal queueing. In fact, such a
2948 	 * device reaches a high speed through internal parallelism
2949 	 * and pipelining. This means that, to reach a high
2950 	 * throughput, it must have many requests enqueued at the same
2951 	 * time. But, in this configuration, the internal scheduling
2952 	 * algorithm of the device does exactly the job of queue
2953 	 * merging: it reorders requests so as to obtain as much as
2954 	 * possible a sequential I/O pattern. As a consequence, with
2955 	 * the workload generated by processes doing interleaved I/O,
2956 	 * the throughput reached by the device is likely to be the
2957 	 * same, with and without queue merging.
2958 	 *
2959 	 * Disabling merging also provides a remarkable benefit in
2960 	 * terms of throughput. Merging tends to make many workloads
2961 	 * artificially more uneven, because of shared queues
2962 	 * remaining non empty for incomparably more time than
2963 	 * non-merged queues. This may accentuate workload
2964 	 * asymmetries. For example, if one of the queues in a set of
2965 	 * merged queues has a higher weight than a normal queue, then
2966 	 * the shared queue may inherit such a high weight and, by
2967 	 * staying almost always active, may force BFQ to perform I/O
2968 	 * plugging most of the time. This evidently makes it harder
2969 	 * for BFQ to let the device reach a high throughput.
2970 	 *
2971 	 * Finally, the likely() macro below is not used because one
2972 	 * of the two branches is more likely than the other, but to
2973 	 * have the code path after the following if() executed as
2974 	 * fast as possible for the case of a non rotational device
2975 	 * with queueing. We want it because this is the fastest kind
2976 	 * of device. On the opposite end, the likely() may lengthen
2977 	 * the execution time of BFQ for the case of slower devices
2978 	 * (rotational or at least without queueing). But in this case
2979 	 * the execution time of BFQ matters very little, if not at
2980 	 * all.
2981 	 */
2982 	if (likely(bfqd->nonrot_with_queueing))
2983 		return NULL;
2984 
2985 	/*
2986 	 * Prevent bfqq from being merged if it has been created too
2987 	 * long ago. The idea is that true cooperating processes, and
2988 	 * thus their associated bfq_queues, are supposed to be
2989 	 * created shortly after each other. This is the case, e.g.,
2990 	 * for KVM/QEMU and dump I/O threads. Basing on this
2991 	 * assumption, the following filtering greatly reduces the
2992 	 * probability that two non-cooperating processes, which just
2993 	 * happen to do close I/O for some short time interval, have
2994 	 * their queues merged by mistake.
2995 	 */
2996 	if (bfq_too_late_for_merging(bfqq))
2997 		return NULL;
2998 
2999 	if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
3000 		return NULL;
3001 
3002 	/* If there is only one backlogged queue, don't search. */
3003 	if (bfq_tot_busy_queues(bfqd) == 1)
3004 		return NULL;
3005 
3006 	in_service_bfqq = bfqd->in_service_queue;
3007 
3008 	if (in_service_bfqq && in_service_bfqq != bfqq &&
3009 	    likely(in_service_bfqq != &bfqd->oom_bfqq) &&
3010 	    bfq_rq_close_to_sector(io_struct, request,
3011 				   bfqd->in_serv_last_pos) &&
3012 	    bfqq->entity.parent == in_service_bfqq->entity.parent &&
3013 	    bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
3014 		new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
3015 		if (new_bfqq)
3016 			return new_bfqq;
3017 	}
3018 	/*
3019 	 * Check whether there is a cooperator among currently scheduled
3020 	 * queues. The only thing we need is that the bio/request is not
3021 	 * NULL, as we need it to establish whether a cooperator exists.
3022 	 */
3023 	new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
3024 			bfq_io_struct_pos(io_struct, request));
3025 
3026 	if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
3027 	    bfq_may_be_close_cooperator(bfqq, new_bfqq))
3028 		return bfq_setup_merge(bfqq, new_bfqq);
3029 
3030 	return NULL;
3031 }
3032 
bfq_bfqq_save_state(struct bfq_queue * bfqq)3033 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
3034 {
3035 	struct bfq_io_cq *bic = bfqq->bic;
3036 
3037 	/*
3038 	 * If !bfqq->bic, the queue is already shared or its requests
3039 	 * have already been redirected to a shared queue; both idle window
3040 	 * and weight raising state have already been saved. Do nothing.
3041 	 */
3042 	if (!bic)
3043 		return;
3044 
3045 	bic->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
3046 	bic->saved_inject_limit = bfqq->inject_limit;
3047 	bic->saved_decrease_time_jif = bfqq->decrease_time_jif;
3048 
3049 	bic->saved_weight = bfqq->entity.orig_weight;
3050 	bic->saved_ttime = bfqq->ttime;
3051 	bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
3052 	bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
3053 	bic->saved_io_start_time = bfqq->io_start_time;
3054 	bic->saved_tot_idle_time = bfqq->tot_idle_time;
3055 	bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
3056 	bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
3057 	if (unlikely(bfq_bfqq_just_created(bfqq) &&
3058 		     !bfq_bfqq_in_large_burst(bfqq) &&
3059 		     bfqq->bfqd->low_latency)) {
3060 		/*
3061 		 * bfqq being merged right after being created: bfqq
3062 		 * would have deserved interactive weight raising, but
3063 		 * did not make it to be set in a weight-raised state,
3064 		 * because of this early merge.	Store directly the
3065 		 * weight-raising state that would have been assigned
3066 		 * to bfqq, so that to avoid that bfqq unjustly fails
3067 		 * to enjoy weight raising if split soon.
3068 		 */
3069 		bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
3070 		bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
3071 		bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
3072 		bic->saved_last_wr_start_finish = jiffies;
3073 	} else {
3074 		bic->saved_wr_coeff = bfqq->wr_coeff;
3075 		bic->saved_wr_start_at_switch_to_srt =
3076 			bfqq->wr_start_at_switch_to_srt;
3077 		bic->saved_service_from_wr = bfqq->service_from_wr;
3078 		bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
3079 		bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3080 	}
3081 }
3082 
3083 
3084 static void
bfq_reassign_last_bfqq(struct bfq_queue * cur_bfqq,struct bfq_queue * new_bfqq)3085 bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
3086 {
3087 	if (cur_bfqq->entity.parent &&
3088 	    cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3089 		cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3090 	else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3091 		cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3092 }
3093 
bfq_release_process_ref(struct bfq_data * bfqd,struct bfq_queue * bfqq)3094 void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3095 {
3096 	/*
3097 	 * To prevent bfqq's service guarantees from being violated,
3098 	 * bfqq may be left busy, i.e., queued for service, even if
3099 	 * empty (see comments in __bfq_bfqq_expire() for
3100 	 * details). But, if no process will send requests to bfqq any
3101 	 * longer, then there is no point in keeping bfqq queued for
3102 	 * service. In addition, keeping bfqq queued for service, but
3103 	 * with no process ref any longer, may have caused bfqq to be
3104 	 * freed when dequeued from service. But this is assumed to
3105 	 * never happen.
3106 	 */
3107 	if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3108 	    bfqq != bfqd->in_service_queue)
3109 		bfq_del_bfqq_busy(bfqq, false);
3110 
3111 	bfq_reassign_last_bfqq(bfqq, NULL);
3112 
3113 	bfq_put_queue(bfqq);
3114 }
3115 
3116 static void
bfq_merge_bfqqs(struct bfq_data * bfqd,struct bfq_io_cq * bic,struct bfq_queue * bfqq,struct bfq_queue * new_bfqq)3117 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
3118 		struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
3119 {
3120 	bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3121 		(unsigned long)new_bfqq->pid);
3122 	/* Save weight raising and idle window of the merged queues */
3123 	bfq_bfqq_save_state(bfqq);
3124 	bfq_bfqq_save_state(new_bfqq);
3125 	if (bfq_bfqq_IO_bound(bfqq))
3126 		bfq_mark_bfqq_IO_bound(new_bfqq);
3127 	bfq_clear_bfqq_IO_bound(bfqq);
3128 
3129 	/*
3130 	 * The processes associated with bfqq are cooperators of the
3131 	 * processes associated with new_bfqq. So, if bfqq has a
3132 	 * waker, then assume that all these processes will be happy
3133 	 * to let bfqq's waker freely inject I/O when they have no
3134 	 * I/O.
3135 	 */
3136 	if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3137 	    bfqq->waker_bfqq != new_bfqq) {
3138 		new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3139 		new_bfqq->tentative_waker_bfqq = NULL;
3140 
3141 		/*
3142 		 * If the waker queue disappears, then
3143 		 * new_bfqq->waker_bfqq must be reset. So insert
3144 		 * new_bfqq into the woken_list of the waker. See
3145 		 * bfq_check_waker for details.
3146 		 */
3147 		hlist_add_head(&new_bfqq->woken_list_node,
3148 			       &new_bfqq->waker_bfqq->woken_list);
3149 
3150 	}
3151 
3152 	/*
3153 	 * If bfqq is weight-raised, then let new_bfqq inherit
3154 	 * weight-raising. To reduce false positives, neglect the case
3155 	 * where bfqq has just been created, but has not yet made it
3156 	 * to be weight-raised (which may happen because EQM may merge
3157 	 * bfqq even before bfq_add_request is executed for the first
3158 	 * time for bfqq). Handling this case would however be very
3159 	 * easy, thanks to the flag just_created.
3160 	 */
3161 	if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3162 		new_bfqq->wr_coeff = bfqq->wr_coeff;
3163 		new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3164 		new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3165 		new_bfqq->wr_start_at_switch_to_srt =
3166 			bfqq->wr_start_at_switch_to_srt;
3167 		if (bfq_bfqq_busy(new_bfqq))
3168 			bfqd->wr_busy_queues++;
3169 		new_bfqq->entity.prio_changed = 1;
3170 	}
3171 
3172 	if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3173 		bfqq->wr_coeff = 1;
3174 		bfqq->entity.prio_changed = 1;
3175 		if (bfq_bfqq_busy(bfqq))
3176 			bfqd->wr_busy_queues--;
3177 	}
3178 
3179 	bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3180 		     bfqd->wr_busy_queues);
3181 
3182 	/*
3183 	 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3184 	 */
3185 	bic_set_bfqq(bic, new_bfqq, true);
3186 	bfq_mark_bfqq_coop(new_bfqq);
3187 	/*
3188 	 * new_bfqq now belongs to at least two bics (it is a shared queue):
3189 	 * set new_bfqq->bic to NULL. bfqq either:
3190 	 * - does not belong to any bic any more, and hence bfqq->bic must
3191 	 *   be set to NULL, or
3192 	 * - is a queue whose owning bics have already been redirected to a
3193 	 *   different queue, hence the queue is destined to not belong to
3194 	 *   any bic soon and bfqq->bic is already NULL (therefore the next
3195 	 *   assignment causes no harm).
3196 	 */
3197 	new_bfqq->bic = NULL;
3198 	/*
3199 	 * If the queue is shared, the pid is the pid of one of the associated
3200 	 * processes. Which pid depends on the exact sequence of merge events
3201 	 * the queue underwent. So printing such a pid is useless and confusing
3202 	 * because it reports a random pid between those of the associated
3203 	 * processes.
3204 	 * We mark such a queue with a pid -1, and then print SHARED instead of
3205 	 * a pid in logging messages.
3206 	 */
3207 	new_bfqq->pid = -1;
3208 	bfqq->bic = NULL;
3209 
3210 	bfq_reassign_last_bfqq(bfqq, new_bfqq);
3211 
3212 	bfq_release_process_ref(bfqd, bfqq);
3213 }
3214 
bfq_allow_bio_merge(struct request_queue * q,struct request * rq,struct bio * bio)3215 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3216 				struct bio *bio)
3217 {
3218 	struct bfq_data *bfqd = q->elevator->elevator_data;
3219 	bool is_sync = op_is_sync(bio->bi_opf);
3220 	struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3221 
3222 	/*
3223 	 * Disallow merge of a sync bio into an async request.
3224 	 */
3225 	if (is_sync && !rq_is_sync(rq))
3226 		return false;
3227 
3228 	/*
3229 	 * Lookup the bfqq that this bio will be queued with. Allow
3230 	 * merge only if rq is queued there.
3231 	 */
3232 	if (!bfqq)
3233 		return false;
3234 
3235 	/*
3236 	 * We take advantage of this function to perform an early merge
3237 	 * of the queues of possible cooperating processes.
3238 	 */
3239 	new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3240 	if (new_bfqq) {
3241 		/*
3242 		 * bic still points to bfqq, then it has not yet been
3243 		 * redirected to some other bfq_queue, and a queue
3244 		 * merge between bfqq and new_bfqq can be safely
3245 		 * fulfilled, i.e., bic can be redirected to new_bfqq
3246 		 * and bfqq can be put.
3247 		 */
3248 		bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
3249 				new_bfqq);
3250 		/*
3251 		 * If we get here, bio will be queued into new_queue,
3252 		 * so use new_bfqq to decide whether bio and rq can be
3253 		 * merged.
3254 		 */
3255 		bfqq = new_bfqq;
3256 
3257 		/*
3258 		 * Change also bqfd->bio_bfqq, as
3259 		 * bfqd->bio_bic now points to new_bfqq, and
3260 		 * this function may be invoked again (and then may
3261 		 * use again bqfd->bio_bfqq).
3262 		 */
3263 		bfqd->bio_bfqq = bfqq;
3264 	}
3265 
3266 	return bfqq == RQ_BFQQ(rq);
3267 }
3268 
3269 /*
3270  * Set the maximum time for the in-service queue to consume its
3271  * budget. This prevents seeky processes from lowering the throughput.
3272  * In practice, a time-slice service scheme is used with seeky
3273  * processes.
3274  */
bfq_set_budget_timeout(struct bfq_data * bfqd,struct bfq_queue * bfqq)3275 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3276 				   struct bfq_queue *bfqq)
3277 {
3278 	unsigned int timeout_coeff;
3279 
3280 	if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3281 		timeout_coeff = 1;
3282 	else
3283 		timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3284 
3285 	bfqd->last_budget_start = ktime_get();
3286 
3287 	bfqq->budget_timeout = jiffies +
3288 		bfqd->bfq_timeout * timeout_coeff;
3289 }
3290 
__bfq_set_in_service_queue(struct bfq_data * bfqd,struct bfq_queue * bfqq)3291 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3292 				       struct bfq_queue *bfqq)
3293 {
3294 	if (bfqq) {
3295 		bfq_clear_bfqq_fifo_expire(bfqq);
3296 
3297 		bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3298 
3299 		if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3300 		    bfqq->wr_coeff > 1 &&
3301 		    bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3302 		    time_is_before_jiffies(bfqq->budget_timeout)) {
3303 			/*
3304 			 * For soft real-time queues, move the start
3305 			 * of the weight-raising period forward by the
3306 			 * time the queue has not received any
3307 			 * service. Otherwise, a relatively long
3308 			 * service delay is likely to cause the
3309 			 * weight-raising period of the queue to end,
3310 			 * because of the short duration of the
3311 			 * weight-raising period of a soft real-time
3312 			 * queue.  It is worth noting that this move
3313 			 * is not so dangerous for the other queues,
3314 			 * because soft real-time queues are not
3315 			 * greedy.
3316 			 *
3317 			 * To not add a further variable, we use the
3318 			 * overloaded field budget_timeout to
3319 			 * determine for how long the queue has not
3320 			 * received service, i.e., how much time has
3321 			 * elapsed since the queue expired. However,
3322 			 * this is a little imprecise, because
3323 			 * budget_timeout is set to jiffies if bfqq
3324 			 * not only expires, but also remains with no
3325 			 * request.
3326 			 */
3327 			if (time_after(bfqq->budget_timeout,
3328 				       bfqq->last_wr_start_finish))
3329 				bfqq->last_wr_start_finish +=
3330 					jiffies - bfqq->budget_timeout;
3331 			else
3332 				bfqq->last_wr_start_finish = jiffies;
3333 		}
3334 
3335 		bfq_set_budget_timeout(bfqd, bfqq);
3336 		bfq_log_bfqq(bfqd, bfqq,
3337 			     "set_in_service_queue, cur-budget = %d",
3338 			     bfqq->entity.budget);
3339 	}
3340 
3341 	bfqd->in_service_queue = bfqq;
3342 	bfqd->in_serv_last_pos = 0;
3343 }
3344 
3345 /*
3346  * Get and set a new queue for service.
3347  */
bfq_set_in_service_queue(struct bfq_data * bfqd)3348 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3349 {
3350 	struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3351 
3352 	__bfq_set_in_service_queue(bfqd, bfqq);
3353 	return bfqq;
3354 }
3355 
bfq_arm_slice_timer(struct bfq_data * bfqd)3356 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3357 {
3358 	struct bfq_queue *bfqq = bfqd->in_service_queue;
3359 	u32 sl;
3360 
3361 	bfq_mark_bfqq_wait_request(bfqq);
3362 
3363 	/*
3364 	 * We don't want to idle for seeks, but we do want to allow
3365 	 * fair distribution of slice time for a process doing back-to-back
3366 	 * seeks. So allow a little bit of time for him to submit a new rq.
3367 	 */
3368 	sl = bfqd->bfq_slice_idle;
3369 	/*
3370 	 * Unless the queue is being weight-raised or the scenario is
3371 	 * asymmetric, grant only minimum idle time if the queue
3372 	 * is seeky. A long idling is preserved for a weight-raised
3373 	 * queue, or, more in general, in an asymmetric scenario,
3374 	 * because a long idling is needed for guaranteeing to a queue
3375 	 * its reserved share of the throughput (in particular, it is
3376 	 * needed if the queue has a higher weight than some other
3377 	 * queue).
3378 	 */
3379 	if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3380 	    !bfq_asymmetric_scenario(bfqd, bfqq))
3381 		sl = min_t(u64, sl, BFQ_MIN_TT);
3382 	else if (bfqq->wr_coeff > 1)
3383 		sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3384 
3385 	bfqd->last_idling_start = ktime_get();
3386 	bfqd->last_idling_start_jiffies = jiffies;
3387 
3388 	hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3389 		      HRTIMER_MODE_REL);
3390 	bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3391 }
3392 
3393 /*
3394  * In autotuning mode, max_budget is dynamically recomputed as the
3395  * amount of sectors transferred in timeout at the estimated peak
3396  * rate. This enables BFQ to utilize a full timeslice with a full
3397  * budget, even if the in-service queue is served at peak rate. And
3398  * this maximises throughput with sequential workloads.
3399  */
bfq_calc_max_budget(struct bfq_data * bfqd)3400 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3401 {
3402 	return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3403 		jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3404 }
3405 
3406 /*
3407  * Update parameters related to throughput and responsiveness, as a
3408  * function of the estimated peak rate. See comments on
3409  * bfq_calc_max_budget(), and on the ref_wr_duration array.
3410  */
update_thr_responsiveness_params(struct bfq_data * bfqd)3411 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3412 {
3413 	if (bfqd->bfq_user_max_budget == 0) {
3414 		bfqd->bfq_max_budget =
3415 			bfq_calc_max_budget(bfqd);
3416 		bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3417 	}
3418 }
3419 
bfq_reset_rate_computation(struct bfq_data * bfqd,struct request * rq)3420 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3421 				       struct request *rq)
3422 {
3423 	if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3424 		bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3425 		bfqd->peak_rate_samples = 1;
3426 		bfqd->sequential_samples = 0;
3427 		bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3428 			blk_rq_sectors(rq);
3429 	} else /* no new rq dispatched, just reset the number of samples */
3430 		bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3431 
3432 	bfq_log(bfqd,
3433 		"reset_rate_computation at end, sample %u/%u tot_sects %llu",
3434 		bfqd->peak_rate_samples, bfqd->sequential_samples,
3435 		bfqd->tot_sectors_dispatched);
3436 }
3437 
bfq_update_rate_reset(struct bfq_data * bfqd,struct request * rq)3438 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3439 {
3440 	u32 rate, weight, divisor;
3441 
3442 	/*
3443 	 * For the convergence property to hold (see comments on
3444 	 * bfq_update_peak_rate()) and for the assessment to be
3445 	 * reliable, a minimum number of samples must be present, and
3446 	 * a minimum amount of time must have elapsed. If not so, do
3447 	 * not compute new rate. Just reset parameters, to get ready
3448 	 * for a new evaluation attempt.
3449 	 */
3450 	if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3451 	    bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3452 		goto reset_computation;
3453 
3454 	/*
3455 	 * If a new request completion has occurred after last
3456 	 * dispatch, then, to approximate the rate at which requests
3457 	 * have been served by the device, it is more precise to
3458 	 * extend the observation interval to the last completion.
3459 	 */
3460 	bfqd->delta_from_first =
3461 		max_t(u64, bfqd->delta_from_first,
3462 		      bfqd->last_completion - bfqd->first_dispatch);
3463 
3464 	/*
3465 	 * Rate computed in sects/usec, and not sects/nsec, for
3466 	 * precision issues.
3467 	 */
3468 	rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3469 			div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3470 
3471 	/*
3472 	 * Peak rate not updated if:
3473 	 * - the percentage of sequential dispatches is below 3/4 of the
3474 	 *   total, and rate is below the current estimated peak rate
3475 	 * - rate is unreasonably high (> 20M sectors/sec)
3476 	 */
3477 	if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3478 	     rate <= bfqd->peak_rate) ||
3479 		rate > 20<<BFQ_RATE_SHIFT)
3480 		goto reset_computation;
3481 
3482 	/*
3483 	 * We have to update the peak rate, at last! To this purpose,
3484 	 * we use a low-pass filter. We compute the smoothing constant
3485 	 * of the filter as a function of the 'weight' of the new
3486 	 * measured rate.
3487 	 *
3488 	 * As can be seen in next formulas, we define this weight as a
3489 	 * quantity proportional to how sequential the workload is,
3490 	 * and to how long the observation time interval is.
3491 	 *
3492 	 * The weight runs from 0 to 8. The maximum value of the
3493 	 * weight, 8, yields the minimum value for the smoothing
3494 	 * constant. At this minimum value for the smoothing constant,
3495 	 * the measured rate contributes for half of the next value of
3496 	 * the estimated peak rate.
3497 	 *
3498 	 * So, the first step is to compute the weight as a function
3499 	 * of how sequential the workload is. Note that the weight
3500 	 * cannot reach 9, because bfqd->sequential_samples cannot
3501 	 * become equal to bfqd->peak_rate_samples, which, in its
3502 	 * turn, holds true because bfqd->sequential_samples is not
3503 	 * incremented for the first sample.
3504 	 */
3505 	weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3506 
3507 	/*
3508 	 * Second step: further refine the weight as a function of the
3509 	 * duration of the observation interval.
3510 	 */
3511 	weight = min_t(u32, 8,
3512 		       div_u64(weight * bfqd->delta_from_first,
3513 			       BFQ_RATE_REF_INTERVAL));
3514 
3515 	/*
3516 	 * Divisor ranging from 10, for minimum weight, to 2, for
3517 	 * maximum weight.
3518 	 */
3519 	divisor = 10 - weight;
3520 
3521 	/*
3522 	 * Finally, update peak rate:
3523 	 *
3524 	 * peak_rate = peak_rate * (divisor-1) / divisor  +  rate / divisor
3525 	 */
3526 	bfqd->peak_rate *= divisor-1;
3527 	bfqd->peak_rate /= divisor;
3528 	rate /= divisor; /* smoothing constant alpha = 1/divisor */
3529 
3530 	bfqd->peak_rate += rate;
3531 
3532 	/*
3533 	 * For a very slow device, bfqd->peak_rate can reach 0 (see
3534 	 * the minimum representable values reported in the comments
3535 	 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3536 	 * divisions by zero where bfqd->peak_rate is used as a
3537 	 * divisor.
3538 	 */
3539 	bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3540 
3541 	update_thr_responsiveness_params(bfqd);
3542 
3543 reset_computation:
3544 	bfq_reset_rate_computation(bfqd, rq);
3545 }
3546 
3547 /*
3548  * Update the read/write peak rate (the main quantity used for
3549  * auto-tuning, see update_thr_responsiveness_params()).
3550  *
3551  * It is not trivial to estimate the peak rate (correctly): because of
3552  * the presence of sw and hw queues between the scheduler and the
3553  * device components that finally serve I/O requests, it is hard to
3554  * say exactly when a given dispatched request is served inside the
3555  * device, and for how long. As a consequence, it is hard to know
3556  * precisely at what rate a given set of requests is actually served
3557  * by the device.
3558  *
3559  * On the opposite end, the dispatch time of any request is trivially
3560  * available, and, from this piece of information, the "dispatch rate"
3561  * of requests can be immediately computed. So, the idea in the next
3562  * function is to use what is known, namely request dispatch times
3563  * (plus, when useful, request completion times), to estimate what is
3564  * unknown, namely in-device request service rate.
3565  *
3566  * The main issue is that, because of the above facts, the rate at
3567  * which a certain set of requests is dispatched over a certain time
3568  * interval can vary greatly with respect to the rate at which the
3569  * same requests are then served. But, since the size of any
3570  * intermediate queue is limited, and the service scheme is lossless
3571  * (no request is silently dropped), the following obvious convergence
3572  * property holds: the number of requests dispatched MUST become
3573  * closer and closer to the number of requests completed as the
3574  * observation interval grows. This is the key property used in
3575  * the next function to estimate the peak service rate as a function
3576  * of the observed dispatch rate. The function assumes to be invoked
3577  * on every request dispatch.
3578  */
bfq_update_peak_rate(struct bfq_data * bfqd,struct request * rq)3579 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3580 {
3581 	u64 now_ns = ktime_get_ns();
3582 
3583 	if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3584 		bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3585 			bfqd->peak_rate_samples);
3586 		bfq_reset_rate_computation(bfqd, rq);
3587 		goto update_last_values; /* will add one sample */
3588 	}
3589 
3590 	/*
3591 	 * Device idle for very long: the observation interval lasting
3592 	 * up to this dispatch cannot be a valid observation interval
3593 	 * for computing a new peak rate (similarly to the late-
3594 	 * completion event in bfq_completed_request()). Go to
3595 	 * update_rate_and_reset to have the following three steps
3596 	 * taken:
3597 	 * - close the observation interval at the last (previous)
3598 	 *   request dispatch or completion
3599 	 * - compute rate, if possible, for that observation interval
3600 	 * - start a new observation interval with this dispatch
3601 	 */
3602 	if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3603 	    bfqd->rq_in_driver == 0)
3604 		goto update_rate_and_reset;
3605 
3606 	/* Update sampling information */
3607 	bfqd->peak_rate_samples++;
3608 
3609 	if ((bfqd->rq_in_driver > 0 ||
3610 		now_ns - bfqd->last_completion < BFQ_MIN_TT)
3611 	    && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3612 		bfqd->sequential_samples++;
3613 
3614 	bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3615 
3616 	/* Reset max observed rq size every 32 dispatches */
3617 	if (likely(bfqd->peak_rate_samples % 32))
3618 		bfqd->last_rq_max_size =
3619 			max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3620 	else
3621 		bfqd->last_rq_max_size = blk_rq_sectors(rq);
3622 
3623 	bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3624 
3625 	/* Target observation interval not yet reached, go on sampling */
3626 	if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3627 		goto update_last_values;
3628 
3629 update_rate_and_reset:
3630 	bfq_update_rate_reset(bfqd, rq);
3631 update_last_values:
3632 	bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3633 	if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3634 		bfqd->in_serv_last_pos = bfqd->last_position;
3635 	bfqd->last_dispatch = now_ns;
3636 }
3637 
3638 /*
3639  * Remove request from internal lists.
3640  */
bfq_dispatch_remove(struct request_queue * q,struct request * rq)3641 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3642 {
3643 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
3644 
3645 	/*
3646 	 * For consistency, the next instruction should have been
3647 	 * executed after removing the request from the queue and
3648 	 * dispatching it.  We execute instead this instruction before
3649 	 * bfq_remove_request() (and hence introduce a temporary
3650 	 * inconsistency), for efficiency.  In fact, should this
3651 	 * dispatch occur for a non in-service bfqq, this anticipated
3652 	 * increment prevents two counters related to bfqq->dispatched
3653 	 * from risking to be, first, uselessly decremented, and then
3654 	 * incremented again when the (new) value of bfqq->dispatched
3655 	 * happens to be taken into account.
3656 	 */
3657 	bfqq->dispatched++;
3658 	bfq_update_peak_rate(q->elevator->elevator_data, rq);
3659 
3660 	bfq_remove_request(q, rq);
3661 }
3662 
3663 /*
3664  * There is a case where idling does not have to be performed for
3665  * throughput concerns, but to preserve the throughput share of
3666  * the process associated with bfqq.
3667  *
3668  * To introduce this case, we can note that allowing the drive
3669  * to enqueue more than one request at a time, and hence
3670  * delegating de facto final scheduling decisions to the
3671  * drive's internal scheduler, entails loss of control on the
3672  * actual request service order. In particular, the critical
3673  * situation is when requests from different processes happen
3674  * to be present, at the same time, in the internal queue(s)
3675  * of the drive. In such a situation, the drive, by deciding
3676  * the service order of the internally-queued requests, does
3677  * determine also the actual throughput distribution among
3678  * these processes. But the drive typically has no notion or
3679  * concern about per-process throughput distribution, and
3680  * makes its decisions only on a per-request basis. Therefore,
3681  * the service distribution enforced by the drive's internal
3682  * scheduler is likely to coincide with the desired throughput
3683  * distribution only in a completely symmetric, or favorably
3684  * skewed scenario where:
3685  * (i-a) each of these processes must get the same throughput as
3686  *	 the others,
3687  * (i-b) in case (i-a) does not hold, it holds that the process
3688  *       associated with bfqq must receive a lower or equal
3689  *	 throughput than any of the other processes;
3690  * (ii)  the I/O of each process has the same properties, in
3691  *       terms of locality (sequential or random), direction
3692  *       (reads or writes), request sizes, greediness
3693  *       (from I/O-bound to sporadic), and so on;
3694 
3695  * In fact, in such a scenario, the drive tends to treat the requests
3696  * of each process in about the same way as the requests of the
3697  * others, and thus to provide each of these processes with about the
3698  * same throughput.  This is exactly the desired throughput
3699  * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3700  * even more convenient distribution for (the process associated with)
3701  * bfqq.
3702  *
3703  * In contrast, in any asymmetric or unfavorable scenario, device
3704  * idling (I/O-dispatch plugging) is certainly needed to guarantee
3705  * that bfqq receives its assigned fraction of the device throughput
3706  * (see [1] for details).
3707  *
3708  * The problem is that idling may significantly reduce throughput with
3709  * certain combinations of types of I/O and devices. An important
3710  * example is sync random I/O on flash storage with command
3711  * queueing. So, unless bfqq falls in cases where idling also boosts
3712  * throughput, it is important to check conditions (i-a), i(-b) and
3713  * (ii) accurately, so as to avoid idling when not strictly needed for
3714  * service guarantees.
3715  *
3716  * Unfortunately, it is extremely difficult to thoroughly check
3717  * condition (ii). And, in case there are active groups, it becomes
3718  * very difficult to check conditions (i-a) and (i-b) too.  In fact,
3719  * if there are active groups, then, for conditions (i-a) or (i-b) to
3720  * become false 'indirectly', it is enough that an active group
3721  * contains more active processes or sub-groups than some other active
3722  * group. More precisely, for conditions (i-a) or (i-b) to become
3723  * false because of such a group, it is not even necessary that the
3724  * group is (still) active: it is sufficient that, even if the group
3725  * has become inactive, some of its descendant processes still have
3726  * some request already dispatched but still waiting for
3727  * completion. In fact, requests have still to be guaranteed their
3728  * share of the throughput even after being dispatched. In this
3729  * respect, it is easy to show that, if a group frequently becomes
3730  * inactive while still having in-flight requests, and if, when this
3731  * happens, the group is not considered in the calculation of whether
3732  * the scenario is asymmetric, then the group may fail to be
3733  * guaranteed its fair share of the throughput (basically because
3734  * idling may not be performed for the descendant processes of the
3735  * group, but it had to be).  We address this issue with the following
3736  * bi-modal behavior, implemented in the function
3737  * bfq_asymmetric_scenario().
3738  *
3739  * If there are groups with requests waiting for completion
3740  * (as commented above, some of these groups may even be
3741  * already inactive), then the scenario is tagged as
3742  * asymmetric, conservatively, without checking any of the
3743  * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3744  * This behavior matches also the fact that groups are created
3745  * exactly if controlling I/O is a primary concern (to
3746  * preserve bandwidth and latency guarantees).
3747  *
3748  * On the opposite end, if there are no groups with requests waiting
3749  * for completion, then only conditions (i-a) and (i-b) are actually
3750  * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3751  * idling is not performed, regardless of whether condition (ii)
3752  * holds.  In other words, only if conditions (i-a) and (i-b) do not
3753  * hold, then idling is allowed, and the device tends to be prevented
3754  * from queueing many requests, possibly of several processes. Since
3755  * there are no groups with requests waiting for completion, then, to
3756  * control conditions (i-a) and (i-b) it is enough to check just
3757  * whether all the queues with requests waiting for completion also
3758  * have the same weight.
3759  *
3760  * Not checking condition (ii) evidently exposes bfqq to the
3761  * risk of getting less throughput than its fair share.
3762  * However, for queues with the same weight, a further
3763  * mechanism, preemption, mitigates or even eliminates this
3764  * problem. And it does so without consequences on overall
3765  * throughput. This mechanism and its benefits are explained
3766  * in the next three paragraphs.
3767  *
3768  * Even if a queue, say Q, is expired when it remains idle, Q
3769  * can still preempt the new in-service queue if the next
3770  * request of Q arrives soon (see the comments on
3771  * bfq_bfqq_update_budg_for_activation). If all queues and
3772  * groups have the same weight, this form of preemption,
3773  * combined with the hole-recovery heuristic described in the
3774  * comments on function bfq_bfqq_update_budg_for_activation,
3775  * are enough to preserve a correct bandwidth distribution in
3776  * the mid term, even without idling. In fact, even if not
3777  * idling allows the internal queues of the device to contain
3778  * many requests, and thus to reorder requests, we can rather
3779  * safely assume that the internal scheduler still preserves a
3780  * minimum of mid-term fairness.
3781  *
3782  * More precisely, this preemption-based, idleless approach
3783  * provides fairness in terms of IOPS, and not sectors per
3784  * second. This can be seen with a simple example. Suppose
3785  * that there are two queues with the same weight, but that
3786  * the first queue receives requests of 8 sectors, while the
3787  * second queue receives requests of 1024 sectors. In
3788  * addition, suppose that each of the two queues contains at
3789  * most one request at a time, which implies that each queue
3790  * always remains idle after it is served. Finally, after
3791  * remaining idle, each queue receives very quickly a new
3792  * request. It follows that the two queues are served
3793  * alternatively, preempting each other if needed. This
3794  * implies that, although both queues have the same weight,
3795  * the queue with large requests receives a service that is
3796  * 1024/8 times as high as the service received by the other
3797  * queue.
3798  *
3799  * The motivation for using preemption instead of idling (for
3800  * queues with the same weight) is that, by not idling,
3801  * service guarantees are preserved (completely or at least in
3802  * part) without minimally sacrificing throughput. And, if
3803  * there is no active group, then the primary expectation for
3804  * this device is probably a high throughput.
3805  *
3806  * We are now left only with explaining the two sub-conditions in the
3807  * additional compound condition that is checked below for deciding
3808  * whether the scenario is asymmetric. To explain the first
3809  * sub-condition, we need to add that the function
3810  * bfq_asymmetric_scenario checks the weights of only
3811  * non-weight-raised queues, for efficiency reasons (see comments on
3812  * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3813  * is checked explicitly here. More precisely, the compound condition
3814  * below takes into account also the fact that, even if bfqq is being
3815  * weight-raised, the scenario is still symmetric if all queues with
3816  * requests waiting for completion happen to be
3817  * weight-raised. Actually, we should be even more precise here, and
3818  * differentiate between interactive weight raising and soft real-time
3819  * weight raising.
3820  *
3821  * The second sub-condition checked in the compound condition is
3822  * whether there is a fair amount of already in-flight I/O not
3823  * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3824  * following reason. The drive may decide to serve in-flight
3825  * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3826  * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3827  * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3828  * basically uncontrolled amount of I/O from other queues may be
3829  * dispatched too, possibly causing the service of bfqq's I/O to be
3830  * delayed even longer in the drive. This problem gets more and more
3831  * serious as the speed and the queue depth of the drive grow,
3832  * because, as these two quantities grow, the probability to find no
3833  * queue busy but many requests in flight grows too. By contrast,
3834  * plugging I/O dispatching minimizes the delay induced by already
3835  * in-flight I/O, and enables bfqq to recover the bandwidth it may
3836  * lose because of this delay.
3837  *
3838  * As a side note, it is worth considering that the above
3839  * device-idling countermeasures may however fail in the following
3840  * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3841  * in a time period during which all symmetry sub-conditions hold, and
3842  * therefore the device is allowed to enqueue many requests, but at
3843  * some later point in time some sub-condition stops to hold, then it
3844  * may become impossible to make requests be served in the desired
3845  * order until all the requests already queued in the device have been
3846  * served. The last sub-condition commented above somewhat mitigates
3847  * this problem for weight-raised queues.
3848  *
3849  * However, as an additional mitigation for this problem, we preserve
3850  * plugging for a special symmetric case that may suddenly turn into
3851  * asymmetric: the case where only bfqq is busy. In this case, not
3852  * expiring bfqq does not cause any harm to any other queues in terms
3853  * of service guarantees. In contrast, it avoids the following unlucky
3854  * sequence of events: (1) bfqq is expired, (2) a new queue with a
3855  * lower weight than bfqq becomes busy (or more queues), (3) the new
3856  * queue is served until a new request arrives for bfqq, (4) when bfqq
3857  * is finally served, there are so many requests of the new queue in
3858  * the drive that the pending requests for bfqq take a lot of time to
3859  * be served. In particular, event (2) may case even already
3860  * dispatched requests of bfqq to be delayed, inside the drive. So, to
3861  * avoid this series of events, the scenario is preventively declared
3862  * as asymmetric also if bfqq is the only busy queues
3863  */
idling_needed_for_service_guarantees(struct bfq_data * bfqd,struct bfq_queue * bfqq)3864 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3865 						 struct bfq_queue *bfqq)
3866 {
3867 	int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3868 
3869 	/* No point in idling for bfqq if it won't get requests any longer */
3870 	if (unlikely(!bfqq_process_refs(bfqq)))
3871 		return false;
3872 
3873 	return (bfqq->wr_coeff > 1 &&
3874 		(bfqd->wr_busy_queues <
3875 		 tot_busy_queues ||
3876 		 bfqd->rq_in_driver >=
3877 		 bfqq->dispatched + 4)) ||
3878 		bfq_asymmetric_scenario(bfqd, bfqq) ||
3879 		tot_busy_queues == 1;
3880 }
3881 
__bfq_bfqq_expire(struct bfq_data * bfqd,struct bfq_queue * bfqq,enum bfqq_expiration reason)3882 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3883 			      enum bfqq_expiration reason)
3884 {
3885 	/*
3886 	 * If this bfqq is shared between multiple processes, check
3887 	 * to make sure that those processes are still issuing I/Os
3888 	 * within the mean seek distance. If not, it may be time to
3889 	 * break the queues apart again.
3890 	 */
3891 	if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3892 		bfq_mark_bfqq_split_coop(bfqq);
3893 
3894 	/*
3895 	 * Consider queues with a higher finish virtual time than
3896 	 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3897 	 * true, then bfqq's bandwidth would be violated if an
3898 	 * uncontrolled amount of I/O from these queues were
3899 	 * dispatched while bfqq is waiting for its new I/O to
3900 	 * arrive. This is exactly what may happen if this is a forced
3901 	 * expiration caused by a preemption attempt, and if bfqq is
3902 	 * not re-scheduled. To prevent this from happening, re-queue
3903 	 * bfqq if it needs I/O-dispatch plugging, even if it is
3904 	 * empty. By doing so, bfqq is granted to be served before the
3905 	 * above queues (provided that bfqq is of course eligible).
3906 	 */
3907 	if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3908 	    !(reason == BFQQE_PREEMPTED &&
3909 	      idling_needed_for_service_guarantees(bfqd, bfqq))) {
3910 		if (bfqq->dispatched == 0)
3911 			/*
3912 			 * Overloading budget_timeout field to store
3913 			 * the time at which the queue remains with no
3914 			 * backlog and no outstanding request; used by
3915 			 * the weight-raising mechanism.
3916 			 */
3917 			bfqq->budget_timeout = jiffies;
3918 
3919 		bfq_del_bfqq_busy(bfqq, true);
3920 	} else {
3921 		bfq_requeue_bfqq(bfqd, bfqq, true);
3922 		/*
3923 		 * Resort priority tree of potential close cooperators.
3924 		 * See comments on bfq_pos_tree_add_move() for the unlikely().
3925 		 */
3926 		if (unlikely(!bfqd->nonrot_with_queueing &&
3927 			     !RB_EMPTY_ROOT(&bfqq->sort_list)))
3928 			bfq_pos_tree_add_move(bfqd, bfqq);
3929 	}
3930 
3931 	/*
3932 	 * All in-service entities must have been properly deactivated
3933 	 * or requeued before executing the next function, which
3934 	 * resets all in-service entities as no more in service. This
3935 	 * may cause bfqq to be freed. If this happens, the next
3936 	 * function returns true.
3937 	 */
3938 	return __bfq_bfqd_reset_in_service(bfqd);
3939 }
3940 
3941 /**
3942  * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3943  * @bfqd: device data.
3944  * @bfqq: queue to update.
3945  * @reason: reason for expiration.
3946  *
3947  * Handle the feedback on @bfqq budget at queue expiration.
3948  * See the body for detailed comments.
3949  */
__bfq_bfqq_recalc_budget(struct bfq_data * bfqd,struct bfq_queue * bfqq,enum bfqq_expiration reason)3950 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3951 				     struct bfq_queue *bfqq,
3952 				     enum bfqq_expiration reason)
3953 {
3954 	struct request *next_rq;
3955 	int budget, min_budget;
3956 
3957 	min_budget = bfq_min_budget(bfqd);
3958 
3959 	if (bfqq->wr_coeff == 1)
3960 		budget = bfqq->max_budget;
3961 	else /*
3962 	      * Use a constant, low budget for weight-raised queues,
3963 	      * to help achieve a low latency. Keep it slightly higher
3964 	      * than the minimum possible budget, to cause a little
3965 	      * bit fewer expirations.
3966 	      */
3967 		budget = 2 * min_budget;
3968 
3969 	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3970 		bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3971 	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3972 		budget, bfq_min_budget(bfqd));
3973 	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3974 		bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3975 
3976 	if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3977 		switch (reason) {
3978 		/*
3979 		 * Caveat: in all the following cases we trade latency
3980 		 * for throughput.
3981 		 */
3982 		case BFQQE_TOO_IDLE:
3983 			/*
3984 			 * This is the only case where we may reduce
3985 			 * the budget: if there is no request of the
3986 			 * process still waiting for completion, then
3987 			 * we assume (tentatively) that the timer has
3988 			 * expired because the batch of requests of
3989 			 * the process could have been served with a
3990 			 * smaller budget.  Hence, betting that
3991 			 * process will behave in the same way when it
3992 			 * becomes backlogged again, we reduce its
3993 			 * next budget.  As long as we guess right,
3994 			 * this budget cut reduces the latency
3995 			 * experienced by the process.
3996 			 *
3997 			 * However, if there are still outstanding
3998 			 * requests, then the process may have not yet
3999 			 * issued its next request just because it is
4000 			 * still waiting for the completion of some of
4001 			 * the still outstanding ones.  So in this
4002 			 * subcase we do not reduce its budget, on the
4003 			 * contrary we increase it to possibly boost
4004 			 * the throughput, as discussed in the
4005 			 * comments to the BUDGET_TIMEOUT case.
4006 			 */
4007 			if (bfqq->dispatched > 0) /* still outstanding reqs */
4008 				budget = min(budget * 2, bfqd->bfq_max_budget);
4009 			else {
4010 				if (budget > 5 * min_budget)
4011 					budget -= 4 * min_budget;
4012 				else
4013 					budget = min_budget;
4014 			}
4015 			break;
4016 		case BFQQE_BUDGET_TIMEOUT:
4017 			/*
4018 			 * We double the budget here because it gives
4019 			 * the chance to boost the throughput if this
4020 			 * is not a seeky process (and has bumped into
4021 			 * this timeout because of, e.g., ZBR).
4022 			 */
4023 			budget = min(budget * 2, bfqd->bfq_max_budget);
4024 			break;
4025 		case BFQQE_BUDGET_EXHAUSTED:
4026 			/*
4027 			 * The process still has backlog, and did not
4028 			 * let either the budget timeout or the disk
4029 			 * idling timeout expire. Hence it is not
4030 			 * seeky, has a short thinktime and may be
4031 			 * happy with a higher budget too. So
4032 			 * definitely increase the budget of this good
4033 			 * candidate to boost the disk throughput.
4034 			 */
4035 			budget = min(budget * 4, bfqd->bfq_max_budget);
4036 			break;
4037 		case BFQQE_NO_MORE_REQUESTS:
4038 			/*
4039 			 * For queues that expire for this reason, it
4040 			 * is particularly important to keep the
4041 			 * budget close to the actual service they
4042 			 * need. Doing so reduces the timestamp
4043 			 * misalignment problem described in the
4044 			 * comments in the body of
4045 			 * __bfq_activate_entity. In fact, suppose
4046 			 * that a queue systematically expires for
4047 			 * BFQQE_NO_MORE_REQUESTS and presents a
4048 			 * new request in time to enjoy timestamp
4049 			 * back-shifting. The larger the budget of the
4050 			 * queue is with respect to the service the
4051 			 * queue actually requests in each service
4052 			 * slot, the more times the queue can be
4053 			 * reactivated with the same virtual finish
4054 			 * time. It follows that, even if this finish
4055 			 * time is pushed to the system virtual time
4056 			 * to reduce the consequent timestamp
4057 			 * misalignment, the queue unjustly enjoys for
4058 			 * many re-activations a lower finish time
4059 			 * than all newly activated queues.
4060 			 *
4061 			 * The service needed by bfqq is measured
4062 			 * quite precisely by bfqq->entity.service.
4063 			 * Since bfqq does not enjoy device idling,
4064 			 * bfqq->entity.service is equal to the number
4065 			 * of sectors that the process associated with
4066 			 * bfqq requested to read/write before waiting
4067 			 * for request completions, or blocking for
4068 			 * other reasons.
4069 			 */
4070 			budget = max_t(int, bfqq->entity.service, min_budget);
4071 			break;
4072 		default:
4073 			return;
4074 		}
4075 	} else if (!bfq_bfqq_sync(bfqq)) {
4076 		/*
4077 		 * Async queues get always the maximum possible
4078 		 * budget, as for them we do not care about latency
4079 		 * (in addition, their ability to dispatch is limited
4080 		 * by the charging factor).
4081 		 */
4082 		budget = bfqd->bfq_max_budget;
4083 	}
4084 
4085 	bfqq->max_budget = budget;
4086 
4087 	if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4088 	    !bfqd->bfq_user_max_budget)
4089 		bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4090 
4091 	/*
4092 	 * If there is still backlog, then assign a new budget, making
4093 	 * sure that it is large enough for the next request.  Since
4094 	 * the finish time of bfqq must be kept in sync with the
4095 	 * budget, be sure to call __bfq_bfqq_expire() *after* this
4096 	 * update.
4097 	 *
4098 	 * If there is no backlog, then no need to update the budget;
4099 	 * it will be updated on the arrival of a new request.
4100 	 */
4101 	next_rq = bfqq->next_rq;
4102 	if (next_rq)
4103 		bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4104 					    bfq_serv_to_charge(next_rq, bfqq));
4105 
4106 	bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4107 			next_rq ? blk_rq_sectors(next_rq) : 0,
4108 			bfqq->entity.budget);
4109 }
4110 
4111 /*
4112  * Return true if the process associated with bfqq is "slow". The slow
4113  * flag is used, in addition to the budget timeout, to reduce the
4114  * amount of service provided to seeky processes, and thus reduce
4115  * their chances to lower the throughput. More details in the comments
4116  * on the function bfq_bfqq_expire().
4117  *
4118  * An important observation is in order: as discussed in the comments
4119  * on the function bfq_update_peak_rate(), with devices with internal
4120  * queues, it is hard if ever possible to know when and for how long
4121  * an I/O request is processed by the device (apart from the trivial
4122  * I/O pattern where a new request is dispatched only after the
4123  * previous one has been completed). This makes it hard to evaluate
4124  * the real rate at which the I/O requests of each bfq_queue are
4125  * served.  In fact, for an I/O scheduler like BFQ, serving a
4126  * bfq_queue means just dispatching its requests during its service
4127  * slot (i.e., until the budget of the queue is exhausted, or the
4128  * queue remains idle, or, finally, a timeout fires). But, during the
4129  * service slot of a bfq_queue, around 100 ms at most, the device may
4130  * be even still processing requests of bfq_queues served in previous
4131  * service slots. On the opposite end, the requests of the in-service
4132  * bfq_queue may be completed after the service slot of the queue
4133  * finishes.
4134  *
4135  * Anyway, unless more sophisticated solutions are used
4136  * (where possible), the sum of the sizes of the requests dispatched
4137  * during the service slot of a bfq_queue is probably the only
4138  * approximation available for the service received by the bfq_queue
4139  * during its service slot. And this sum is the quantity used in this
4140  * function to evaluate the I/O speed of a process.
4141  */
bfq_bfqq_is_slow(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool compensate,enum bfqq_expiration reason,unsigned long * delta_ms)4142 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4143 				 bool compensate, enum bfqq_expiration reason,
4144 				 unsigned long *delta_ms)
4145 {
4146 	ktime_t delta_ktime;
4147 	u32 delta_usecs;
4148 	bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4149 
4150 	if (!bfq_bfqq_sync(bfqq))
4151 		return false;
4152 
4153 	if (compensate)
4154 		delta_ktime = bfqd->last_idling_start;
4155 	else
4156 		delta_ktime = ktime_get();
4157 	delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4158 	delta_usecs = ktime_to_us(delta_ktime);
4159 
4160 	/* don't use too short time intervals */
4161 	if (delta_usecs < 1000) {
4162 		if (blk_queue_nonrot(bfqd->queue))
4163 			 /*
4164 			  * give same worst-case guarantees as idling
4165 			  * for seeky
4166 			  */
4167 			*delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4168 		else /* charge at least one seek */
4169 			*delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4170 
4171 		return slow;
4172 	}
4173 
4174 	*delta_ms = delta_usecs / USEC_PER_MSEC;
4175 
4176 	/*
4177 	 * Use only long (> 20ms) intervals to filter out excessive
4178 	 * spikes in service rate estimation.
4179 	 */
4180 	if (delta_usecs > 20000) {
4181 		/*
4182 		 * Caveat for rotational devices: processes doing I/O
4183 		 * in the slower disk zones tend to be slow(er) even
4184 		 * if not seeky. In this respect, the estimated peak
4185 		 * rate is likely to be an average over the disk
4186 		 * surface. Accordingly, to not be too harsh with
4187 		 * unlucky processes, a process is deemed slow only if
4188 		 * its rate has been lower than half of the estimated
4189 		 * peak rate.
4190 		 */
4191 		slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4192 	}
4193 
4194 	bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4195 
4196 	return slow;
4197 }
4198 
4199 /*
4200  * To be deemed as soft real-time, an application must meet two
4201  * requirements. First, the application must not require an average
4202  * bandwidth higher than the approximate bandwidth required to playback or
4203  * record a compressed high-definition video.
4204  * The next function is invoked on the completion of the last request of a
4205  * batch, to compute the next-start time instant, soft_rt_next_start, such
4206  * that, if the next request of the application does not arrive before
4207  * soft_rt_next_start, then the above requirement on the bandwidth is met.
4208  *
4209  * The second requirement is that the request pattern of the application is
4210  * isochronous, i.e., that, after issuing a request or a batch of requests,
4211  * the application stops issuing new requests until all its pending requests
4212  * have been completed. After that, the application may issue a new batch,
4213  * and so on.
4214  * For this reason the next function is invoked to compute
4215  * soft_rt_next_start only for applications that meet this requirement,
4216  * whereas soft_rt_next_start is set to infinity for applications that do
4217  * not.
4218  *
4219  * Unfortunately, even a greedy (i.e., I/O-bound) application may
4220  * happen to meet, occasionally or systematically, both the above
4221  * bandwidth and isochrony requirements. This may happen at least in
4222  * the following circumstances. First, if the CPU load is high. The
4223  * application may stop issuing requests while the CPUs are busy
4224  * serving other processes, then restart, then stop again for a while,
4225  * and so on. The other circumstances are related to the storage
4226  * device: the storage device is highly loaded or reaches a low-enough
4227  * throughput with the I/O of the application (e.g., because the I/O
4228  * is random and/or the device is slow). In all these cases, the
4229  * I/O of the application may be simply slowed down enough to meet
4230  * the bandwidth and isochrony requirements. To reduce the probability
4231  * that greedy applications are deemed as soft real-time in these
4232  * corner cases, a further rule is used in the computation of
4233  * soft_rt_next_start: the return value of this function is forced to
4234  * be higher than the maximum between the following two quantities.
4235  *
4236  * (a) Current time plus: (1) the maximum time for which the arrival
4237  *     of a request is waited for when a sync queue becomes idle,
4238  *     namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4239  *     postpone for a moment the reason for adding a few extra
4240  *     jiffies; we get back to it after next item (b).  Lower-bounding
4241  *     the return value of this function with the current time plus
4242  *     bfqd->bfq_slice_idle tends to filter out greedy applications,
4243  *     because the latter issue their next request as soon as possible
4244  *     after the last one has been completed. In contrast, a soft
4245  *     real-time application spends some time processing data, after a
4246  *     batch of its requests has been completed.
4247  *
4248  * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4249  *     above, greedy applications may happen to meet both the
4250  *     bandwidth and isochrony requirements under heavy CPU or
4251  *     storage-device load. In more detail, in these scenarios, these
4252  *     applications happen, only for limited time periods, to do I/O
4253  *     slowly enough to meet all the requirements described so far,
4254  *     including the filtering in above item (a). These slow-speed
4255  *     time intervals are usually interspersed between other time
4256  *     intervals during which these applications do I/O at a very high
4257  *     speed. Fortunately, exactly because of the high speed of the
4258  *     I/O in the high-speed intervals, the values returned by this
4259  *     function happen to be so high, near the end of any such
4260  *     high-speed interval, to be likely to fall *after* the end of
4261  *     the low-speed time interval that follows. These high values are
4262  *     stored in bfqq->soft_rt_next_start after each invocation of
4263  *     this function. As a consequence, if the last value of
4264  *     bfqq->soft_rt_next_start is constantly used to lower-bound the
4265  *     next value that this function may return, then, from the very
4266  *     beginning of a low-speed interval, bfqq->soft_rt_next_start is
4267  *     likely to be constantly kept so high that any I/O request
4268  *     issued during the low-speed interval is considered as arriving
4269  *     to soon for the application to be deemed as soft
4270  *     real-time. Then, in the high-speed interval that follows, the
4271  *     application will not be deemed as soft real-time, just because
4272  *     it will do I/O at a high speed. And so on.
4273  *
4274  * Getting back to the filtering in item (a), in the following two
4275  * cases this filtering might be easily passed by a greedy
4276  * application, if the reference quantity was just
4277  * bfqd->bfq_slice_idle:
4278  * 1) HZ is so low that the duration of a jiffy is comparable to or
4279  *    higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4280  *    devices with HZ=100. The time granularity may be so coarse
4281  *    that the approximation, in jiffies, of bfqd->bfq_slice_idle
4282  *    is rather lower than the exact value.
4283  * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4284  *    for a while, then suddenly 'jump' by several units to recover the lost
4285  *    increments. This seems to happen, e.g., inside virtual machines.
4286  * To address this issue, in the filtering in (a) we do not use as a
4287  * reference time interval just bfqd->bfq_slice_idle, but
4288  * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4289  * minimum number of jiffies for which the filter seems to be quite
4290  * precise also in embedded systems and KVM/QEMU virtual machines.
4291  */
bfq_bfqq_softrt_next_start(struct bfq_data * bfqd,struct bfq_queue * bfqq)4292 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4293 						struct bfq_queue *bfqq)
4294 {
4295 	return max3(bfqq->soft_rt_next_start,
4296 		    bfqq->last_idle_bklogged +
4297 		    HZ * bfqq->service_from_backlogged /
4298 		    bfqd->bfq_wr_max_softrt_rate,
4299 		    jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4300 }
4301 
4302 /**
4303  * bfq_bfqq_expire - expire a queue.
4304  * @bfqd: device owning the queue.
4305  * @bfqq: the queue to expire.
4306  * @compensate: if true, compensate for the time spent idling.
4307  * @reason: the reason causing the expiration.
4308  *
4309  * If the process associated with bfqq does slow I/O (e.g., because it
4310  * issues random requests), we charge bfqq with the time it has been
4311  * in service instead of the service it has received (see
4312  * bfq_bfqq_charge_time for details on how this goal is achieved). As
4313  * a consequence, bfqq will typically get higher timestamps upon
4314  * reactivation, and hence it will be rescheduled as if it had
4315  * received more service than what it has actually received. In the
4316  * end, bfqq receives less service in proportion to how slowly its
4317  * associated process consumes its budgets (and hence how seriously it
4318  * tends to lower the throughput). In addition, this time-charging
4319  * strategy guarantees time fairness among slow processes. In
4320  * contrast, if the process associated with bfqq is not slow, we
4321  * charge bfqq exactly with the service it has received.
4322  *
4323  * Charging time to the first type of queues and the exact service to
4324  * the other has the effect of using the WF2Q+ policy to schedule the
4325  * former on a timeslice basis, without violating service domain
4326  * guarantees among the latter.
4327  */
bfq_bfqq_expire(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool compensate,enum bfqq_expiration reason)4328 void bfq_bfqq_expire(struct bfq_data *bfqd,
4329 		     struct bfq_queue *bfqq,
4330 		     bool compensate,
4331 		     enum bfqq_expiration reason)
4332 {
4333 	bool slow;
4334 	unsigned long delta = 0;
4335 	struct bfq_entity *entity = &bfqq->entity;
4336 
4337 	/*
4338 	 * Check whether the process is slow (see bfq_bfqq_is_slow).
4339 	 */
4340 	slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
4341 
4342 	/*
4343 	 * As above explained, charge slow (typically seeky) and
4344 	 * timed-out queues with the time and not the service
4345 	 * received, to favor sequential workloads.
4346 	 *
4347 	 * Processes doing I/O in the slower disk zones will tend to
4348 	 * be slow(er) even if not seeky. Therefore, since the
4349 	 * estimated peak rate is actually an average over the disk
4350 	 * surface, these processes may timeout just for bad luck. To
4351 	 * avoid punishing them, do not charge time to processes that
4352 	 * succeeded in consuming at least 2/3 of their budget. This
4353 	 * allows BFQ to preserve enough elasticity to still perform
4354 	 * bandwidth, and not time, distribution with little unlucky
4355 	 * or quasi-sequential processes.
4356 	 */
4357 	if (bfqq->wr_coeff == 1 &&
4358 	    (slow ||
4359 	     (reason == BFQQE_BUDGET_TIMEOUT &&
4360 	      bfq_bfqq_budget_left(bfqq) >=  entity->budget / 3)))
4361 		bfq_bfqq_charge_time(bfqd, bfqq, delta);
4362 
4363 	if (bfqd->low_latency && bfqq->wr_coeff == 1)
4364 		bfqq->last_wr_start_finish = jiffies;
4365 
4366 	if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4367 	    RB_EMPTY_ROOT(&bfqq->sort_list)) {
4368 		/*
4369 		 * If we get here, and there are no outstanding
4370 		 * requests, then the request pattern is isochronous
4371 		 * (see the comments on the function
4372 		 * bfq_bfqq_softrt_next_start()). Therefore we can
4373 		 * compute soft_rt_next_start.
4374 		 *
4375 		 * If, instead, the queue still has outstanding
4376 		 * requests, then we have to wait for the completion
4377 		 * of all the outstanding requests to discover whether
4378 		 * the request pattern is actually isochronous.
4379 		 */
4380 		if (bfqq->dispatched == 0)
4381 			bfqq->soft_rt_next_start =
4382 				bfq_bfqq_softrt_next_start(bfqd, bfqq);
4383 		else if (bfqq->dispatched > 0) {
4384 			/*
4385 			 * Schedule an update of soft_rt_next_start to when
4386 			 * the task may be discovered to be isochronous.
4387 			 */
4388 			bfq_mark_bfqq_softrt_update(bfqq);
4389 		}
4390 	}
4391 
4392 	bfq_log_bfqq(bfqd, bfqq,
4393 		"expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4394 		slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4395 
4396 	/*
4397 	 * bfqq expired, so no total service time needs to be computed
4398 	 * any longer: reset state machine for measuring total service
4399 	 * times.
4400 	 */
4401 	bfqd->rqs_injected = bfqd->wait_dispatch = false;
4402 	bfqd->waited_rq = NULL;
4403 
4404 	/*
4405 	 * Increase, decrease or leave budget unchanged according to
4406 	 * reason.
4407 	 */
4408 	__bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4409 	if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4410 		/* bfqq is gone, no more actions on it */
4411 		return;
4412 
4413 	/* mark bfqq as waiting a request only if a bic still points to it */
4414 	if (!bfq_bfqq_busy(bfqq) &&
4415 	    reason != BFQQE_BUDGET_TIMEOUT &&
4416 	    reason != BFQQE_BUDGET_EXHAUSTED) {
4417 		bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4418 		/*
4419 		 * Not setting service to 0, because, if the next rq
4420 		 * arrives in time, the queue will go on receiving
4421 		 * service with this same budget (as if it never expired)
4422 		 */
4423 	} else
4424 		entity->service = 0;
4425 
4426 	/*
4427 	 * Reset the received-service counter for every parent entity.
4428 	 * Differently from what happens with bfqq->entity.service,
4429 	 * the resetting of this counter never needs to be postponed
4430 	 * for parent entities. In fact, in case bfqq may have a
4431 	 * chance to go on being served using the last, partially
4432 	 * consumed budget, bfqq->entity.service needs to be kept,
4433 	 * because if bfqq then actually goes on being served using
4434 	 * the same budget, the last value of bfqq->entity.service is
4435 	 * needed to properly decrement bfqq->entity.budget by the
4436 	 * portion already consumed. In contrast, it is not necessary
4437 	 * to keep entity->service for parent entities too, because
4438 	 * the bubble up of the new value of bfqq->entity.budget will
4439 	 * make sure that the budgets of parent entities are correct,
4440 	 * even in case bfqq and thus parent entities go on receiving
4441 	 * service with the same budget.
4442 	 */
4443 	entity = entity->parent;
4444 	for_each_entity(entity)
4445 		entity->service = 0;
4446 }
4447 
4448 /*
4449  * Budget timeout is not implemented through a dedicated timer, but
4450  * just checked on request arrivals and completions, as well as on
4451  * idle timer expirations.
4452  */
bfq_bfqq_budget_timeout(struct bfq_queue * bfqq)4453 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4454 {
4455 	return time_is_before_eq_jiffies(bfqq->budget_timeout);
4456 }
4457 
4458 /*
4459  * If we expire a queue that is actively waiting (i.e., with the
4460  * device idled) for the arrival of a new request, then we may incur
4461  * the timestamp misalignment problem described in the body of the
4462  * function __bfq_activate_entity. Hence we return true only if this
4463  * condition does not hold, or if the queue is slow enough to deserve
4464  * only to be kicked off for preserving a high throughput.
4465  */
bfq_may_expire_for_budg_timeout(struct bfq_queue * bfqq)4466 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4467 {
4468 	bfq_log_bfqq(bfqq->bfqd, bfqq,
4469 		"may_budget_timeout: wait_request %d left %d timeout %d",
4470 		bfq_bfqq_wait_request(bfqq),
4471 			bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3,
4472 		bfq_bfqq_budget_timeout(bfqq));
4473 
4474 	return (!bfq_bfqq_wait_request(bfqq) ||
4475 		bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3)
4476 		&&
4477 		bfq_bfqq_budget_timeout(bfqq);
4478 }
4479 
idling_boosts_thr_without_issues(struct bfq_data * bfqd,struct bfq_queue * bfqq)4480 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4481 					     struct bfq_queue *bfqq)
4482 {
4483 	bool rot_without_queueing =
4484 		!blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4485 		bfqq_sequential_and_IO_bound,
4486 		idling_boosts_thr;
4487 
4488 	/* No point in idling for bfqq if it won't get requests any longer */
4489 	if (unlikely(!bfqq_process_refs(bfqq)))
4490 		return false;
4491 
4492 	bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4493 		bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4494 
4495 	/*
4496 	 * The next variable takes into account the cases where idling
4497 	 * boosts the throughput.
4498 	 *
4499 	 * The value of the variable is computed considering, first, that
4500 	 * idling is virtually always beneficial for the throughput if:
4501 	 * (a) the device is not NCQ-capable and rotational, or
4502 	 * (b) regardless of the presence of NCQ, the device is rotational and
4503 	 *     the request pattern for bfqq is I/O-bound and sequential, or
4504 	 * (c) regardless of whether it is rotational, the device is
4505 	 *     not NCQ-capable and the request pattern for bfqq is
4506 	 *     I/O-bound and sequential.
4507 	 *
4508 	 * Secondly, and in contrast to the above item (b), idling an
4509 	 * NCQ-capable flash-based device would not boost the
4510 	 * throughput even with sequential I/O; rather it would lower
4511 	 * the throughput in proportion to how fast the device
4512 	 * is. Accordingly, the next variable is true if any of the
4513 	 * above conditions (a), (b) or (c) is true, and, in
4514 	 * particular, happens to be false if bfqd is an NCQ-capable
4515 	 * flash-based device.
4516 	 */
4517 	idling_boosts_thr = rot_without_queueing ||
4518 		((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4519 		 bfqq_sequential_and_IO_bound);
4520 
4521 	/*
4522 	 * The return value of this function is equal to that of
4523 	 * idling_boosts_thr, unless a special case holds. In this
4524 	 * special case, described below, idling may cause problems to
4525 	 * weight-raised queues.
4526 	 *
4527 	 * When the request pool is saturated (e.g., in the presence
4528 	 * of write hogs), if the processes associated with
4529 	 * non-weight-raised queues ask for requests at a lower rate,
4530 	 * then processes associated with weight-raised queues have a
4531 	 * higher probability to get a request from the pool
4532 	 * immediately (or at least soon) when they need one. Thus
4533 	 * they have a higher probability to actually get a fraction
4534 	 * of the device throughput proportional to their high
4535 	 * weight. This is especially true with NCQ-capable drives,
4536 	 * which enqueue several requests in advance, and further
4537 	 * reorder internally-queued requests.
4538 	 *
4539 	 * For this reason, we force to false the return value if
4540 	 * there are weight-raised busy queues. In this case, and if
4541 	 * bfqq is not weight-raised, this guarantees that the device
4542 	 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4543 	 * then idling will be guaranteed by another variable, see
4544 	 * below). Combined with the timestamping rules of BFQ (see
4545 	 * [1] for details), this behavior causes bfqq, and hence any
4546 	 * sync non-weight-raised queue, to get a lower number of
4547 	 * requests served, and thus to ask for a lower number of
4548 	 * requests from the request pool, before the busy
4549 	 * weight-raised queues get served again. This often mitigates
4550 	 * starvation problems in the presence of heavy write
4551 	 * workloads and NCQ, thereby guaranteeing a higher
4552 	 * application and system responsiveness in these hostile
4553 	 * scenarios.
4554 	 */
4555 	return idling_boosts_thr &&
4556 		bfqd->wr_busy_queues == 0;
4557 }
4558 
4559 /*
4560  * For a queue that becomes empty, device idling is allowed only if
4561  * this function returns true for that queue. As a consequence, since
4562  * device idling plays a critical role for both throughput boosting
4563  * and service guarantees, the return value of this function plays a
4564  * critical role as well.
4565  *
4566  * In a nutshell, this function returns true only if idling is
4567  * beneficial for throughput or, even if detrimental for throughput,
4568  * idling is however necessary to preserve service guarantees (low
4569  * latency, desired throughput distribution, ...). In particular, on
4570  * NCQ-capable devices, this function tries to return false, so as to
4571  * help keep the drives' internal queues full, whenever this helps the
4572  * device boost the throughput without causing any service-guarantee
4573  * issue.
4574  *
4575  * Most of the issues taken into account to get the return value of
4576  * this function are not trivial. We discuss these issues in the two
4577  * functions providing the main pieces of information needed by this
4578  * function.
4579  */
bfq_better_to_idle(struct bfq_queue * bfqq)4580 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4581 {
4582 	struct bfq_data *bfqd = bfqq->bfqd;
4583 	bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4584 
4585 	/* No point in idling for bfqq if it won't get requests any longer */
4586 	if (unlikely(!bfqq_process_refs(bfqq)))
4587 		return false;
4588 
4589 	if (unlikely(bfqd->strict_guarantees))
4590 		return true;
4591 
4592 	/*
4593 	 * Idling is performed only if slice_idle > 0. In addition, we
4594 	 * do not idle if
4595 	 * (a) bfqq is async
4596 	 * (b) bfqq is in the idle io prio class: in this case we do
4597 	 * not idle because we want to minimize the bandwidth that
4598 	 * queues in this class can steal to higher-priority queues
4599 	 */
4600 	if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4601 	   bfq_class_idle(bfqq))
4602 		return false;
4603 
4604 	idling_boosts_thr_with_no_issue =
4605 		idling_boosts_thr_without_issues(bfqd, bfqq);
4606 
4607 	idling_needed_for_service_guar =
4608 		idling_needed_for_service_guarantees(bfqd, bfqq);
4609 
4610 	/*
4611 	 * We have now the two components we need to compute the
4612 	 * return value of the function, which is true only if idling
4613 	 * either boosts the throughput (without issues), or is
4614 	 * necessary to preserve service guarantees.
4615 	 */
4616 	return idling_boosts_thr_with_no_issue ||
4617 		idling_needed_for_service_guar;
4618 }
4619 
4620 /*
4621  * If the in-service queue is empty but the function bfq_better_to_idle
4622  * returns true, then:
4623  * 1) the queue must remain in service and cannot be expired, and
4624  * 2) the device must be idled to wait for the possible arrival of a new
4625  *    request for the queue.
4626  * See the comments on the function bfq_better_to_idle for the reasons
4627  * why performing device idling is the best choice to boost the throughput
4628  * and preserve service guarantees when bfq_better_to_idle itself
4629  * returns true.
4630  */
bfq_bfqq_must_idle(struct bfq_queue * bfqq)4631 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4632 {
4633 	return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4634 }
4635 
4636 /*
4637  * This function chooses the queue from which to pick the next extra
4638  * I/O request to inject, if it finds a compatible queue. See the
4639  * comments on bfq_update_inject_limit() for details on the injection
4640  * mechanism, and for the definitions of the quantities mentioned
4641  * below.
4642  */
4643 static struct bfq_queue *
bfq_choose_bfqq_for_injection(struct bfq_data * bfqd)4644 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4645 {
4646 	struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4647 	unsigned int limit = in_serv_bfqq->inject_limit;
4648 	/*
4649 	 * If
4650 	 * - bfqq is not weight-raised and therefore does not carry
4651 	 *   time-critical I/O,
4652 	 * or
4653 	 * - regardless of whether bfqq is weight-raised, bfqq has
4654 	 *   however a long think time, during which it can absorb the
4655 	 *   effect of an appropriate number of extra I/O requests
4656 	 *   from other queues (see bfq_update_inject_limit for
4657 	 *   details on the computation of this number);
4658 	 * then injection can be performed without restrictions.
4659 	 */
4660 	bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4661 		!bfq_bfqq_has_short_ttime(in_serv_bfqq);
4662 
4663 	/*
4664 	 * If
4665 	 * - the baseline total service time could not be sampled yet,
4666 	 *   so the inject limit happens to be still 0, and
4667 	 * - a lot of time has elapsed since the plugging of I/O
4668 	 *   dispatching started, so drive speed is being wasted
4669 	 *   significantly;
4670 	 * then temporarily raise inject limit to one request.
4671 	 */
4672 	if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4673 	    bfq_bfqq_wait_request(in_serv_bfqq) &&
4674 	    time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4675 				      bfqd->bfq_slice_idle)
4676 		)
4677 		limit = 1;
4678 
4679 	if (bfqd->rq_in_driver >= limit)
4680 		return NULL;
4681 
4682 	/*
4683 	 * Linear search of the source queue for injection; but, with
4684 	 * a high probability, very few steps are needed to find a
4685 	 * candidate queue, i.e., a queue with enough budget left for
4686 	 * its next request. In fact:
4687 	 * - BFQ dynamically updates the budget of every queue so as
4688 	 *   to accommodate the expected backlog of the queue;
4689 	 * - if a queue gets all its requests dispatched as injected
4690 	 *   service, then the queue is removed from the active list
4691 	 *   (and re-added only if it gets new requests, but then it
4692 	 *   is assigned again enough budget for its new backlog).
4693 	 */
4694 	list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4695 		if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4696 		    (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4697 		    bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4698 		    bfq_bfqq_budget_left(bfqq)) {
4699 			/*
4700 			 * Allow for only one large in-flight request
4701 			 * on non-rotational devices, for the
4702 			 * following reason. On non-rotationl drives,
4703 			 * large requests take much longer than
4704 			 * smaller requests to be served. In addition,
4705 			 * the drive prefers to serve large requests
4706 			 * w.r.t. to small ones, if it can choose. So,
4707 			 * having more than one large requests queued
4708 			 * in the drive may easily make the next first
4709 			 * request of the in-service queue wait for so
4710 			 * long to break bfqq's service guarantees. On
4711 			 * the bright side, large requests let the
4712 			 * drive reach a very high throughput, even if
4713 			 * there is only one in-flight large request
4714 			 * at a time.
4715 			 */
4716 			if (blk_queue_nonrot(bfqd->queue) &&
4717 			    blk_rq_sectors(bfqq->next_rq) >=
4718 			    BFQQ_SECT_THR_NONROT)
4719 				limit = min_t(unsigned int, 1, limit);
4720 			else
4721 				limit = in_serv_bfqq->inject_limit;
4722 
4723 			if (bfqd->rq_in_driver < limit) {
4724 				bfqd->rqs_injected = true;
4725 				return bfqq;
4726 			}
4727 		}
4728 
4729 	return NULL;
4730 }
4731 
4732 /*
4733  * Select a queue for service.  If we have a current queue in service,
4734  * check whether to continue servicing it, or retrieve and set a new one.
4735  */
bfq_select_queue(struct bfq_data * bfqd)4736 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4737 {
4738 	struct bfq_queue *bfqq;
4739 	struct request *next_rq;
4740 	enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4741 
4742 	bfqq = bfqd->in_service_queue;
4743 	if (!bfqq)
4744 		goto new_queue;
4745 
4746 	bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4747 
4748 	/*
4749 	 * Do not expire bfqq for budget timeout if bfqq may be about
4750 	 * to enjoy device idling. The reason why, in this case, we
4751 	 * prevent bfqq from expiring is the same as in the comments
4752 	 * on the case where bfq_bfqq_must_idle() returns true, in
4753 	 * bfq_completed_request().
4754 	 */
4755 	if (bfq_may_expire_for_budg_timeout(bfqq) &&
4756 	    !bfq_bfqq_must_idle(bfqq))
4757 		goto expire;
4758 
4759 check_queue:
4760 	/*
4761 	 * This loop is rarely executed more than once. Even when it
4762 	 * happens, it is much more convenient to re-execute this loop
4763 	 * than to return NULL and trigger a new dispatch to get a
4764 	 * request served.
4765 	 */
4766 	next_rq = bfqq->next_rq;
4767 	/*
4768 	 * If bfqq has requests queued and it has enough budget left to
4769 	 * serve them, keep the queue, otherwise expire it.
4770 	 */
4771 	if (next_rq) {
4772 		if (bfq_serv_to_charge(next_rq, bfqq) >
4773 			bfq_bfqq_budget_left(bfqq)) {
4774 			/*
4775 			 * Expire the queue for budget exhaustion,
4776 			 * which makes sure that the next budget is
4777 			 * enough to serve the next request, even if
4778 			 * it comes from the fifo expired path.
4779 			 */
4780 			reason = BFQQE_BUDGET_EXHAUSTED;
4781 			goto expire;
4782 		} else {
4783 			/*
4784 			 * The idle timer may be pending because we may
4785 			 * not disable disk idling even when a new request
4786 			 * arrives.
4787 			 */
4788 			if (bfq_bfqq_wait_request(bfqq)) {
4789 				/*
4790 				 * If we get here: 1) at least a new request
4791 				 * has arrived but we have not disabled the
4792 				 * timer because the request was too small,
4793 				 * 2) then the block layer has unplugged
4794 				 * the device, causing the dispatch to be
4795 				 * invoked.
4796 				 *
4797 				 * Since the device is unplugged, now the
4798 				 * requests are probably large enough to
4799 				 * provide a reasonable throughput.
4800 				 * So we disable idling.
4801 				 */
4802 				bfq_clear_bfqq_wait_request(bfqq);
4803 				hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4804 			}
4805 			goto keep_queue;
4806 		}
4807 	}
4808 
4809 	/*
4810 	 * No requests pending. However, if the in-service queue is idling
4811 	 * for a new request, or has requests waiting for a completion and
4812 	 * may idle after their completion, then keep it anyway.
4813 	 *
4814 	 * Yet, inject service from other queues if it boosts
4815 	 * throughput and is possible.
4816 	 */
4817 	if (bfq_bfqq_wait_request(bfqq) ||
4818 	    (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4819 		struct bfq_queue *async_bfqq =
4820 			bfqq->bic && bfqq->bic->bfqq[0] &&
4821 			bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4822 			bfqq->bic->bfqq[0]->next_rq ?
4823 			bfqq->bic->bfqq[0] : NULL;
4824 		struct bfq_queue *blocked_bfqq =
4825 			!hlist_empty(&bfqq->woken_list) ?
4826 			container_of(bfqq->woken_list.first,
4827 				     struct bfq_queue,
4828 				     woken_list_node)
4829 			: NULL;
4830 
4831 		/*
4832 		 * The next four mutually-exclusive ifs decide
4833 		 * whether to try injection, and choose the queue to
4834 		 * pick an I/O request from.
4835 		 *
4836 		 * The first if checks whether the process associated
4837 		 * with bfqq has also async I/O pending. If so, it
4838 		 * injects such I/O unconditionally. Injecting async
4839 		 * I/O from the same process can cause no harm to the
4840 		 * process. On the contrary, it can only increase
4841 		 * bandwidth and reduce latency for the process.
4842 		 *
4843 		 * The second if checks whether there happens to be a
4844 		 * non-empty waker queue for bfqq, i.e., a queue whose
4845 		 * I/O needs to be completed for bfqq to receive new
4846 		 * I/O. This happens, e.g., if bfqq is associated with
4847 		 * a process that does some sync. A sync generates
4848 		 * extra blocking I/O, which must be completed before
4849 		 * the process associated with bfqq can go on with its
4850 		 * I/O. If the I/O of the waker queue is not served,
4851 		 * then bfqq remains empty, and no I/O is dispatched,
4852 		 * until the idle timeout fires for bfqq. This is
4853 		 * likely to result in lower bandwidth and higher
4854 		 * latencies for bfqq, and in a severe loss of total
4855 		 * throughput. The best action to take is therefore to
4856 		 * serve the waker queue as soon as possible. So do it
4857 		 * (without relying on the third alternative below for
4858 		 * eventually serving waker_bfqq's I/O; see the last
4859 		 * paragraph for further details). This systematic
4860 		 * injection of I/O from the waker queue does not
4861 		 * cause any delay to bfqq's I/O. On the contrary,
4862 		 * next bfqq's I/O is brought forward dramatically,
4863 		 * for it is not blocked for milliseconds.
4864 		 *
4865 		 * The third if checks whether there is a queue woken
4866 		 * by bfqq, and currently with pending I/O. Such a
4867 		 * woken queue does not steal bandwidth from bfqq,
4868 		 * because it remains soon without I/O if bfqq is not
4869 		 * served. So there is virtually no risk of loss of
4870 		 * bandwidth for bfqq if this woken queue has I/O
4871 		 * dispatched while bfqq is waiting for new I/O.
4872 		 *
4873 		 * The fourth if checks whether bfqq is a queue for
4874 		 * which it is better to avoid injection. It is so if
4875 		 * bfqq delivers more throughput when served without
4876 		 * any further I/O from other queues in the middle, or
4877 		 * if the service times of bfqq's I/O requests both
4878 		 * count more than overall throughput, and may be
4879 		 * easily increased by injection (this happens if bfqq
4880 		 * has a short think time). If none of these
4881 		 * conditions holds, then a candidate queue for
4882 		 * injection is looked for through
4883 		 * bfq_choose_bfqq_for_injection(). Note that the
4884 		 * latter may return NULL (for example if the inject
4885 		 * limit for bfqq is currently 0).
4886 		 *
4887 		 * NOTE: motivation for the second alternative
4888 		 *
4889 		 * Thanks to the way the inject limit is updated in
4890 		 * bfq_update_has_short_ttime(), it is rather likely
4891 		 * that, if I/O is being plugged for bfqq and the
4892 		 * waker queue has pending I/O requests that are
4893 		 * blocking bfqq's I/O, then the fourth alternative
4894 		 * above lets the waker queue get served before the
4895 		 * I/O-plugging timeout fires. So one may deem the
4896 		 * second alternative superfluous. It is not, because
4897 		 * the fourth alternative may be way less effective in
4898 		 * case of a synchronization. For two main
4899 		 * reasons. First, throughput may be low because the
4900 		 * inject limit may be too low to guarantee the same
4901 		 * amount of injected I/O, from the waker queue or
4902 		 * other queues, that the second alternative
4903 		 * guarantees (the second alternative unconditionally
4904 		 * injects a pending I/O request of the waker queue
4905 		 * for each bfq_dispatch_request()). Second, with the
4906 		 * fourth alternative, the duration of the plugging,
4907 		 * i.e., the time before bfqq finally receives new I/O,
4908 		 * may not be minimized, because the waker queue may
4909 		 * happen to be served only after other queues.
4910 		 */
4911 		if (async_bfqq &&
4912 		    icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4913 		    bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4914 		    bfq_bfqq_budget_left(async_bfqq))
4915 			bfqq = bfqq->bic->bfqq[0];
4916 		else if (bfqq->waker_bfqq &&
4917 			   bfq_bfqq_busy(bfqq->waker_bfqq) &&
4918 			   bfqq->waker_bfqq->next_rq &&
4919 			   bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4920 					      bfqq->waker_bfqq) <=
4921 			   bfq_bfqq_budget_left(bfqq->waker_bfqq)
4922 			)
4923 			bfqq = bfqq->waker_bfqq;
4924 		else if (blocked_bfqq &&
4925 			   bfq_bfqq_busy(blocked_bfqq) &&
4926 			   blocked_bfqq->next_rq &&
4927 			   bfq_serv_to_charge(blocked_bfqq->next_rq,
4928 					      blocked_bfqq) <=
4929 			   bfq_bfqq_budget_left(blocked_bfqq)
4930 			)
4931 			bfqq = blocked_bfqq;
4932 		else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4933 			 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4934 			  !bfq_bfqq_has_short_ttime(bfqq)))
4935 			bfqq = bfq_choose_bfqq_for_injection(bfqd);
4936 		else
4937 			bfqq = NULL;
4938 
4939 		goto keep_queue;
4940 	}
4941 
4942 	reason = BFQQE_NO_MORE_REQUESTS;
4943 expire:
4944 	bfq_bfqq_expire(bfqd, bfqq, false, reason);
4945 new_queue:
4946 	bfqq = bfq_set_in_service_queue(bfqd);
4947 	if (bfqq) {
4948 		bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4949 		goto check_queue;
4950 	}
4951 keep_queue:
4952 	if (bfqq)
4953 		bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4954 	else
4955 		bfq_log(bfqd, "select_queue: no queue returned");
4956 
4957 	return bfqq;
4958 }
4959 
bfq_update_wr_data(struct bfq_data * bfqd,struct bfq_queue * bfqq)4960 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4961 {
4962 	struct bfq_entity *entity = &bfqq->entity;
4963 
4964 	if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4965 		bfq_log_bfqq(bfqd, bfqq,
4966 			"raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4967 			jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4968 			jiffies_to_msecs(bfqq->wr_cur_max_time),
4969 			bfqq->wr_coeff,
4970 			bfqq->entity.weight, bfqq->entity.orig_weight);
4971 
4972 		if (entity->prio_changed)
4973 			bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4974 
4975 		/*
4976 		 * If the queue was activated in a burst, or too much
4977 		 * time has elapsed from the beginning of this
4978 		 * weight-raising period, then end weight raising.
4979 		 */
4980 		if (bfq_bfqq_in_large_burst(bfqq))
4981 			bfq_bfqq_end_wr(bfqq);
4982 		else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4983 						bfqq->wr_cur_max_time)) {
4984 			if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4985 			time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4986 					       bfq_wr_duration(bfqd))) {
4987 				/*
4988 				 * Either in interactive weight
4989 				 * raising, or in soft_rt weight
4990 				 * raising with the
4991 				 * interactive-weight-raising period
4992 				 * elapsed (so no switch back to
4993 				 * interactive weight raising).
4994 				 */
4995 				bfq_bfqq_end_wr(bfqq);
4996 			} else { /*
4997 				  * soft_rt finishing while still in
4998 				  * interactive period, switch back to
4999 				  * interactive weight raising
5000 				  */
5001 				switch_back_to_interactive_wr(bfqq, bfqd);
5002 				bfqq->entity.prio_changed = 1;
5003 			}
5004 		}
5005 		if (bfqq->wr_coeff > 1 &&
5006 		    bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
5007 		    bfqq->service_from_wr > max_service_from_wr) {
5008 			/* see comments on max_service_from_wr */
5009 			bfq_bfqq_end_wr(bfqq);
5010 		}
5011 	}
5012 	/*
5013 	 * To improve latency (for this or other queues), immediately
5014 	 * update weight both if it must be raised and if it must be
5015 	 * lowered. Since, entity may be on some active tree here, and
5016 	 * might have a pending change of its ioprio class, invoke
5017 	 * next function with the last parameter unset (see the
5018 	 * comments on the function).
5019 	 */
5020 	if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
5021 		__bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
5022 						entity, false);
5023 }
5024 
5025 /*
5026  * Dispatch next request from bfqq.
5027  */
bfq_dispatch_rq_from_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq)5028 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
5029 						 struct bfq_queue *bfqq)
5030 {
5031 	struct request *rq = bfqq->next_rq;
5032 	unsigned long service_to_charge;
5033 
5034 	service_to_charge = bfq_serv_to_charge(rq, bfqq);
5035 
5036 	bfq_bfqq_served(bfqq, service_to_charge);
5037 
5038 	if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
5039 		bfqd->wait_dispatch = false;
5040 		bfqd->waited_rq = rq;
5041 	}
5042 
5043 	bfq_dispatch_remove(bfqd->queue, rq);
5044 
5045 	if (bfqq != bfqd->in_service_queue)
5046 		goto return_rq;
5047 
5048 	/*
5049 	 * If weight raising has to terminate for bfqq, then next
5050 	 * function causes an immediate update of bfqq's weight,
5051 	 * without waiting for next activation. As a consequence, on
5052 	 * expiration, bfqq will be timestamped as if has never been
5053 	 * weight-raised during this service slot, even if it has
5054 	 * received part or even most of the service as a
5055 	 * weight-raised queue. This inflates bfqq's timestamps, which
5056 	 * is beneficial, as bfqq is then more willing to leave the
5057 	 * device immediately to possible other weight-raised queues.
5058 	 */
5059 	bfq_update_wr_data(bfqd, bfqq);
5060 
5061 	/*
5062 	 * Expire bfqq, pretending that its budget expired, if bfqq
5063 	 * belongs to CLASS_IDLE and other queues are waiting for
5064 	 * service.
5065 	 */
5066 	if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
5067 		goto return_rq;
5068 
5069 	bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5070 
5071 return_rq:
5072 	return rq;
5073 }
5074 
bfq_has_work(struct blk_mq_hw_ctx * hctx)5075 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5076 {
5077 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5078 
5079 	/*
5080 	 * Avoiding lock: a race on bfqd->queued should cause at
5081 	 * most a call to dispatch for nothing
5082 	 */
5083 	return !list_empty_careful(&bfqd->dispatch) ||
5084 		READ_ONCE(bfqd->queued);
5085 }
5086 
__bfq_dispatch_request(struct blk_mq_hw_ctx * hctx)5087 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5088 {
5089 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5090 	struct request *rq = NULL;
5091 	struct bfq_queue *bfqq = NULL;
5092 
5093 	if (!list_empty(&bfqd->dispatch)) {
5094 		rq = list_first_entry(&bfqd->dispatch, struct request,
5095 				      queuelist);
5096 		list_del_init(&rq->queuelist);
5097 
5098 		bfqq = RQ_BFQQ(rq);
5099 
5100 		if (bfqq) {
5101 			/*
5102 			 * Increment counters here, because this
5103 			 * dispatch does not follow the standard
5104 			 * dispatch flow (where counters are
5105 			 * incremented)
5106 			 */
5107 			bfqq->dispatched++;
5108 
5109 			goto inc_in_driver_start_rq;
5110 		}
5111 
5112 		/*
5113 		 * We exploit the bfq_finish_requeue_request hook to
5114 		 * decrement rq_in_driver, but
5115 		 * bfq_finish_requeue_request will not be invoked on
5116 		 * this request. So, to avoid unbalance, just start
5117 		 * this request, without incrementing rq_in_driver. As
5118 		 * a negative consequence, rq_in_driver is deceptively
5119 		 * lower than it should be while this request is in
5120 		 * service. This may cause bfq_schedule_dispatch to be
5121 		 * invoked uselessly.
5122 		 *
5123 		 * As for implementing an exact solution, the
5124 		 * bfq_finish_requeue_request hook, if defined, is
5125 		 * probably invoked also on this request. So, by
5126 		 * exploiting this hook, we could 1) increment
5127 		 * rq_in_driver here, and 2) decrement it in
5128 		 * bfq_finish_requeue_request. Such a solution would
5129 		 * let the value of the counter be always accurate,
5130 		 * but it would entail using an extra interface
5131 		 * function. This cost seems higher than the benefit,
5132 		 * being the frequency of non-elevator-private
5133 		 * requests very low.
5134 		 */
5135 		goto start_rq;
5136 	}
5137 
5138 	bfq_log(bfqd, "dispatch requests: %d busy queues",
5139 		bfq_tot_busy_queues(bfqd));
5140 
5141 	if (bfq_tot_busy_queues(bfqd) == 0)
5142 		goto exit;
5143 
5144 	/*
5145 	 * Force device to serve one request at a time if
5146 	 * strict_guarantees is true. Forcing this service scheme is
5147 	 * currently the ONLY way to guarantee that the request
5148 	 * service order enforced by the scheduler is respected by a
5149 	 * queueing device. Otherwise the device is free even to make
5150 	 * some unlucky request wait for as long as the device
5151 	 * wishes.
5152 	 *
5153 	 * Of course, serving one request at a time may cause loss of
5154 	 * throughput.
5155 	 */
5156 	if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
5157 		goto exit;
5158 
5159 	bfqq = bfq_select_queue(bfqd);
5160 	if (!bfqq)
5161 		goto exit;
5162 
5163 	rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5164 
5165 	if (rq) {
5166 inc_in_driver_start_rq:
5167 		bfqd->rq_in_driver++;
5168 start_rq:
5169 		rq->rq_flags |= RQF_STARTED;
5170 	}
5171 exit:
5172 	return rq;
5173 }
5174 
5175 #ifdef CONFIG_BFQ_CGROUP_DEBUG
bfq_update_dispatch_stats(struct request_queue * q,struct request * rq,struct bfq_queue * in_serv_queue,bool idle_timer_disabled)5176 static void bfq_update_dispatch_stats(struct request_queue *q,
5177 				      struct request *rq,
5178 				      struct bfq_queue *in_serv_queue,
5179 				      bool idle_timer_disabled)
5180 {
5181 	struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5182 
5183 	if (!idle_timer_disabled && !bfqq)
5184 		return;
5185 
5186 	/*
5187 	 * rq and bfqq are guaranteed to exist until this function
5188 	 * ends, for the following reasons. First, rq can be
5189 	 * dispatched to the device, and then can be completed and
5190 	 * freed, only after this function ends. Second, rq cannot be
5191 	 * merged (and thus freed because of a merge) any longer,
5192 	 * because it has already started. Thus rq cannot be freed
5193 	 * before this function ends, and, since rq has a reference to
5194 	 * bfqq, the same guarantee holds for bfqq too.
5195 	 *
5196 	 * In addition, the following queue lock guarantees that
5197 	 * bfqq_group(bfqq) exists as well.
5198 	 */
5199 	spin_lock_irq(&q->queue_lock);
5200 	if (idle_timer_disabled)
5201 		/*
5202 		 * Since the idle timer has been disabled,
5203 		 * in_serv_queue contained some request when
5204 		 * __bfq_dispatch_request was invoked above, which
5205 		 * implies that rq was picked exactly from
5206 		 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5207 		 * therefore guaranteed to exist because of the above
5208 		 * arguments.
5209 		 */
5210 		bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5211 	if (bfqq) {
5212 		struct bfq_group *bfqg = bfqq_group(bfqq);
5213 
5214 		bfqg_stats_update_avg_queue_size(bfqg);
5215 		bfqg_stats_set_start_empty_time(bfqg);
5216 		bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5217 	}
5218 	spin_unlock_irq(&q->queue_lock);
5219 }
5220 #else
bfq_update_dispatch_stats(struct request_queue * q,struct request * rq,struct bfq_queue * in_serv_queue,bool idle_timer_disabled)5221 static inline void bfq_update_dispatch_stats(struct request_queue *q,
5222 					     struct request *rq,
5223 					     struct bfq_queue *in_serv_queue,
5224 					     bool idle_timer_disabled) {}
5225 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5226 
bfq_dispatch_request(struct blk_mq_hw_ctx * hctx)5227 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5228 {
5229 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5230 	struct request *rq;
5231 	struct bfq_queue *in_serv_queue;
5232 	bool waiting_rq, idle_timer_disabled = false;
5233 
5234 	spin_lock_irq(&bfqd->lock);
5235 
5236 	in_serv_queue = bfqd->in_service_queue;
5237 	waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5238 
5239 	rq = __bfq_dispatch_request(hctx);
5240 	if (in_serv_queue == bfqd->in_service_queue) {
5241 		idle_timer_disabled =
5242 			waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5243 	}
5244 
5245 	spin_unlock_irq(&bfqd->lock);
5246 	bfq_update_dispatch_stats(hctx->queue, rq,
5247 			idle_timer_disabled ? in_serv_queue : NULL,
5248 				idle_timer_disabled);
5249 
5250 	return rq;
5251 }
5252 
5253 /*
5254  * Task holds one reference to the queue, dropped when task exits.  Each rq
5255  * in-flight on this queue also holds a reference, dropped when rq is freed.
5256  *
5257  * Scheduler lock must be held here. Recall not to use bfqq after calling
5258  * this function on it.
5259  */
bfq_put_queue(struct bfq_queue * bfqq)5260 void bfq_put_queue(struct bfq_queue *bfqq)
5261 {
5262 	struct bfq_queue *item;
5263 	struct hlist_node *n;
5264 	struct bfq_group *bfqg = bfqq_group(bfqq);
5265 
5266 	bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", bfqq, bfqq->ref);
5267 
5268 	bfqq->ref--;
5269 	if (bfqq->ref)
5270 		return;
5271 
5272 	if (!hlist_unhashed(&bfqq->burst_list_node)) {
5273 		hlist_del_init(&bfqq->burst_list_node);
5274 		/*
5275 		 * Decrement also burst size after the removal, if the
5276 		 * process associated with bfqq is exiting, and thus
5277 		 * does not contribute to the burst any longer. This
5278 		 * decrement helps filter out false positives of large
5279 		 * bursts, when some short-lived process (often due to
5280 		 * the execution of commands by some service) happens
5281 		 * to start and exit while a complex application is
5282 		 * starting, and thus spawning several processes that
5283 		 * do I/O (and that *must not* be treated as a large
5284 		 * burst, see comments on bfq_handle_burst).
5285 		 *
5286 		 * In particular, the decrement is performed only if:
5287 		 * 1) bfqq is not a merged queue, because, if it is,
5288 		 * then this free of bfqq is not triggered by the exit
5289 		 * of the process bfqq is associated with, but exactly
5290 		 * by the fact that bfqq has just been merged.
5291 		 * 2) burst_size is greater than 0, to handle
5292 		 * unbalanced decrements. Unbalanced decrements may
5293 		 * happen in te following case: bfqq is inserted into
5294 		 * the current burst list--without incrementing
5295 		 * bust_size--because of a split, but the current
5296 		 * burst list is not the burst list bfqq belonged to
5297 		 * (see comments on the case of a split in
5298 		 * bfq_set_request).
5299 		 */
5300 		if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5301 			bfqq->bfqd->burst_size--;
5302 	}
5303 
5304 	/*
5305 	 * bfqq does not exist any longer, so it cannot be woken by
5306 	 * any other queue, and cannot wake any other queue. Then bfqq
5307 	 * must be removed from the woken list of its possible waker
5308 	 * queue, and all queues in the woken list of bfqq must stop
5309 	 * having a waker queue. Strictly speaking, these updates
5310 	 * should be performed when bfqq remains with no I/O source
5311 	 * attached to it, which happens before bfqq gets freed. In
5312 	 * particular, this happens when the last process associated
5313 	 * with bfqq exits or gets associated with a different
5314 	 * queue. However, both events lead to bfqq being freed soon,
5315 	 * and dangling references would come out only after bfqq gets
5316 	 * freed. So these updates are done here, as a simple and safe
5317 	 * way to handle all cases.
5318 	 */
5319 	/* remove bfqq from woken list */
5320 	if (!hlist_unhashed(&bfqq->woken_list_node))
5321 		hlist_del_init(&bfqq->woken_list_node);
5322 
5323 	/* reset waker for all queues in woken list */
5324 	hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5325 				  woken_list_node) {
5326 		item->waker_bfqq = NULL;
5327 		hlist_del_init(&item->woken_list_node);
5328 	}
5329 
5330 	if (bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5331 		bfqq->bfqd->last_completed_rq_bfqq = NULL;
5332 
5333 	kmem_cache_free(bfq_pool, bfqq);
5334 	bfqg_and_blkg_put(bfqg);
5335 }
5336 
bfq_put_stable_ref(struct bfq_queue * bfqq)5337 static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5338 {
5339 	bfqq->stable_ref--;
5340 	bfq_put_queue(bfqq);
5341 }
5342 
bfq_put_cooperator(struct bfq_queue * bfqq)5343 void bfq_put_cooperator(struct bfq_queue *bfqq)
5344 {
5345 	struct bfq_queue *__bfqq, *next;
5346 
5347 	/*
5348 	 * If this queue was scheduled to merge with another queue, be
5349 	 * sure to drop the reference taken on that queue (and others in
5350 	 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5351 	 */
5352 	__bfqq = bfqq->new_bfqq;
5353 	while (__bfqq) {
5354 		if (__bfqq == bfqq)
5355 			break;
5356 		next = __bfqq->new_bfqq;
5357 		bfq_put_queue(__bfqq);
5358 		__bfqq = next;
5359 	}
5360 }
5361 
bfq_exit_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq)5362 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5363 {
5364 	if (bfqq == bfqd->in_service_queue) {
5365 		__bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5366 		bfq_schedule_dispatch(bfqd);
5367 	}
5368 
5369 	bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5370 
5371 	bfq_put_cooperator(bfqq);
5372 
5373 	bfq_release_process_ref(bfqd, bfqq);
5374 }
5375 
bfq_exit_icq_bfqq(struct bfq_io_cq * bic,bool is_sync)5376 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
5377 {
5378 	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5379 	struct bfq_data *bfqd;
5380 
5381 	if (bfqq)
5382 		bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5383 
5384 	if (bfqq && bfqd) {
5385 		unsigned long flags;
5386 
5387 		spin_lock_irqsave(&bfqd->lock, flags);
5388 		bic_set_bfqq(bic, NULL, is_sync);
5389 		bfq_exit_bfqq(bfqd, bfqq);
5390 		spin_unlock_irqrestore(&bfqd->lock, flags);
5391 	}
5392 }
5393 
bfq_exit_icq(struct io_cq * icq)5394 static void bfq_exit_icq(struct io_cq *icq)
5395 {
5396 	struct bfq_io_cq *bic = icq_to_bic(icq);
5397 
5398 	if (bic->stable_merge_bfqq) {
5399 		struct bfq_data *bfqd = bic->stable_merge_bfqq->bfqd;
5400 
5401 		/*
5402 		 * bfqd is NULL if scheduler already exited, and in
5403 		 * that case this is the last time bfqq is accessed.
5404 		 */
5405 		if (bfqd) {
5406 			unsigned long flags;
5407 
5408 			spin_lock_irqsave(&bfqd->lock, flags);
5409 			bfq_put_stable_ref(bic->stable_merge_bfqq);
5410 			spin_unlock_irqrestore(&bfqd->lock, flags);
5411 		} else {
5412 			bfq_put_stable_ref(bic->stable_merge_bfqq);
5413 		}
5414 	}
5415 
5416 	bfq_exit_icq_bfqq(bic, true);
5417 	bfq_exit_icq_bfqq(bic, false);
5418 }
5419 
5420 /*
5421  * Update the entity prio values; note that the new values will not
5422  * be used until the next (re)activation.
5423  */
5424 static void
bfq_set_next_ioprio_data(struct bfq_queue * bfqq,struct bfq_io_cq * bic)5425 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5426 {
5427 	struct task_struct *tsk = current;
5428 	int ioprio_class;
5429 	struct bfq_data *bfqd = bfqq->bfqd;
5430 
5431 	if (!bfqd)
5432 		return;
5433 
5434 	ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5435 	switch (ioprio_class) {
5436 	default:
5437 		pr_err("bdi %s: bfq: bad prio class %d\n",
5438 			bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5439 			ioprio_class);
5440 		fallthrough;
5441 	case IOPRIO_CLASS_NONE:
5442 		/*
5443 		 * No prio set, inherit CPU scheduling settings.
5444 		 */
5445 		bfqq->new_ioprio = task_nice_ioprio(tsk);
5446 		bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5447 		break;
5448 	case IOPRIO_CLASS_RT:
5449 		bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5450 		bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5451 		break;
5452 	case IOPRIO_CLASS_BE:
5453 		bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5454 		bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5455 		break;
5456 	case IOPRIO_CLASS_IDLE:
5457 		bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5458 		bfqq->new_ioprio = 7;
5459 		break;
5460 	}
5461 
5462 	if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5463 		pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5464 			bfqq->new_ioprio);
5465 		bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5466 	}
5467 
5468 	bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5469 	bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5470 		     bfqq->new_ioprio, bfqq->entity.new_weight);
5471 	bfqq->entity.prio_changed = 1;
5472 }
5473 
5474 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5475 				       struct bio *bio, bool is_sync,
5476 				       struct bfq_io_cq *bic,
5477 				       bool respawn);
5478 
bfq_check_ioprio_change(struct bfq_io_cq * bic,struct bio * bio)5479 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5480 {
5481 	struct bfq_data *bfqd = bic_to_bfqd(bic);
5482 	struct bfq_queue *bfqq;
5483 	int ioprio = bic->icq.ioc->ioprio;
5484 
5485 	/*
5486 	 * This condition may trigger on a newly created bic, be sure to
5487 	 * drop the lock before returning.
5488 	 */
5489 	if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5490 		return;
5491 
5492 	bic->ioprio = ioprio;
5493 
5494 	bfqq = bic_to_bfqq(bic, false);
5495 	if (bfqq) {
5496 		struct bfq_queue *old_bfqq = bfqq;
5497 
5498 		bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5499 		bic_set_bfqq(bic, bfqq, false);
5500 		bfq_release_process_ref(bfqd, old_bfqq);
5501 	}
5502 
5503 	bfqq = bic_to_bfqq(bic, true);
5504 	if (bfqq)
5505 		bfq_set_next_ioprio_data(bfqq, bic);
5506 }
5507 
bfq_init_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct bfq_io_cq * bic,pid_t pid,int is_sync)5508 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5509 			  struct bfq_io_cq *bic, pid_t pid, int is_sync)
5510 {
5511 	u64 now_ns = ktime_get_ns();
5512 
5513 	RB_CLEAR_NODE(&bfqq->entity.rb_node);
5514 	INIT_LIST_HEAD(&bfqq->fifo);
5515 	INIT_HLIST_NODE(&bfqq->burst_list_node);
5516 	INIT_HLIST_NODE(&bfqq->woken_list_node);
5517 	INIT_HLIST_HEAD(&bfqq->woken_list);
5518 
5519 	bfqq->ref = 0;
5520 	bfqq->bfqd = bfqd;
5521 
5522 	if (bic)
5523 		bfq_set_next_ioprio_data(bfqq, bic);
5524 
5525 	if (is_sync) {
5526 		/*
5527 		 * No need to mark as has_short_ttime if in
5528 		 * idle_class, because no device idling is performed
5529 		 * for queues in idle class
5530 		 */
5531 		if (!bfq_class_idle(bfqq))
5532 			/* tentatively mark as has_short_ttime */
5533 			bfq_mark_bfqq_has_short_ttime(bfqq);
5534 		bfq_mark_bfqq_sync(bfqq);
5535 		bfq_mark_bfqq_just_created(bfqq);
5536 	} else
5537 		bfq_clear_bfqq_sync(bfqq);
5538 
5539 	/* set end request to minus infinity from now */
5540 	bfqq->ttime.last_end_request = now_ns + 1;
5541 
5542 	bfqq->creation_time = jiffies;
5543 
5544 	bfqq->io_start_time = now_ns;
5545 
5546 	bfq_mark_bfqq_IO_bound(bfqq);
5547 
5548 	bfqq->pid = pid;
5549 
5550 	/* Tentative initial value to trade off between thr and lat */
5551 	bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5552 	bfqq->budget_timeout = bfq_smallest_from_now();
5553 
5554 	bfqq->wr_coeff = 1;
5555 	bfqq->last_wr_start_finish = jiffies;
5556 	bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5557 	bfqq->split_time = bfq_smallest_from_now();
5558 
5559 	/*
5560 	 * To not forget the possibly high bandwidth consumed by a
5561 	 * process/queue in the recent past,
5562 	 * bfq_bfqq_softrt_next_start() returns a value at least equal
5563 	 * to the current value of bfqq->soft_rt_next_start (see
5564 	 * comments on bfq_bfqq_softrt_next_start).  Set
5565 	 * soft_rt_next_start to now, to mean that bfqq has consumed
5566 	 * no bandwidth so far.
5567 	 */
5568 	bfqq->soft_rt_next_start = jiffies;
5569 
5570 	/* first request is almost certainly seeky */
5571 	bfqq->seek_history = 1;
5572 }
5573 
bfq_async_queue_prio(struct bfq_data * bfqd,struct bfq_group * bfqg,int ioprio_class,int ioprio)5574 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5575 					       struct bfq_group *bfqg,
5576 					       int ioprio_class, int ioprio)
5577 {
5578 	switch (ioprio_class) {
5579 	case IOPRIO_CLASS_RT:
5580 		return &bfqg->async_bfqq[0][ioprio];
5581 	case IOPRIO_CLASS_NONE:
5582 		ioprio = IOPRIO_BE_NORM;
5583 		fallthrough;
5584 	case IOPRIO_CLASS_BE:
5585 		return &bfqg->async_bfqq[1][ioprio];
5586 	case IOPRIO_CLASS_IDLE:
5587 		return &bfqg->async_idle_bfqq;
5588 	default:
5589 		return NULL;
5590 	}
5591 }
5592 
5593 static struct bfq_queue *
bfq_do_early_stable_merge(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct bfq_io_cq * bic,struct bfq_queue * last_bfqq_created)5594 bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5595 			  struct bfq_io_cq *bic,
5596 			  struct bfq_queue *last_bfqq_created)
5597 {
5598 	struct bfq_queue *new_bfqq =
5599 		bfq_setup_merge(bfqq, last_bfqq_created);
5600 
5601 	if (!new_bfqq)
5602 		return bfqq;
5603 
5604 	if (new_bfqq->bic)
5605 		new_bfqq->bic->stably_merged = true;
5606 	bic->stably_merged = true;
5607 
5608 	/*
5609 	 * Reusing merge functions. This implies that
5610 	 * bfqq->bic must be set too, for
5611 	 * bfq_merge_bfqqs to correctly save bfqq's
5612 	 * state before killing it.
5613 	 */
5614 	bfqq->bic = bic;
5615 	bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
5616 
5617 	return new_bfqq;
5618 }
5619 
5620 /*
5621  * Many throughput-sensitive workloads are made of several parallel
5622  * I/O flows, with all flows generated by the same application, or
5623  * more generically by the same task (e.g., system boot). The most
5624  * counterproductive action with these workloads is plugging I/O
5625  * dispatch when one of the bfq_queues associated with these flows
5626  * remains temporarily empty.
5627  *
5628  * To avoid this plugging, BFQ has been using a burst-handling
5629  * mechanism for years now. This mechanism has proven effective for
5630  * throughput, and not detrimental for service guarantees. The
5631  * following function pushes this mechanism a little bit further,
5632  * basing on the following two facts.
5633  *
5634  * First, all the I/O flows of a the same application or task
5635  * contribute to the execution/completion of that common application
5636  * or task. So the performance figures that matter are total
5637  * throughput of the flows and task-wide I/O latency.  In particular,
5638  * these flows do not need to be protected from each other, in terms
5639  * of individual bandwidth or latency.
5640  *
5641  * Second, the above fact holds regardless of the number of flows.
5642  *
5643  * Putting these two facts together, this commits merges stably the
5644  * bfq_queues associated with these I/O flows, i.e., with the
5645  * processes that generate these IO/ flows, regardless of how many the
5646  * involved processes are.
5647  *
5648  * To decide whether a set of bfq_queues is actually associated with
5649  * the I/O flows of a common application or task, and to merge these
5650  * queues stably, this function operates as follows: given a bfq_queue,
5651  * say Q2, currently being created, and the last bfq_queue, say Q1,
5652  * created before Q2, Q2 is merged stably with Q1 if
5653  * - very little time has elapsed since when Q1 was created
5654  * - Q2 has the same ioprio as Q1
5655  * - Q2 belongs to the same group as Q1
5656  *
5657  * Merging bfq_queues also reduces scheduling overhead. A fio test
5658  * with ten random readers on /dev/nullb shows a throughput boost of
5659  * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5660  * the total per-request processing time, the above throughput boost
5661  * implies that BFQ's overhead is reduced by more than 50%.
5662  *
5663  * This new mechanism most certainly obsoletes the current
5664  * burst-handling heuristics. We keep those heuristics for the moment.
5665  */
bfq_do_or_sched_stable_merge(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct bfq_io_cq * bic)5666 static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5667 						      struct bfq_queue *bfqq,
5668 						      struct bfq_io_cq *bic)
5669 {
5670 	struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5671 		&bfqq->entity.parent->last_bfqq_created :
5672 		&bfqd->last_bfqq_created;
5673 
5674 	struct bfq_queue *last_bfqq_created = *source_bfqq;
5675 
5676 	/*
5677 	 * If last_bfqq_created has not been set yet, then init it. If
5678 	 * it has been set already, but too long ago, then move it
5679 	 * forward to bfqq. Finally, move also if bfqq belongs to a
5680 	 * different group than last_bfqq_created, or if bfqq has a
5681 	 * different ioprio or ioprio_class. If none of these
5682 	 * conditions holds true, then try an early stable merge or
5683 	 * schedule a delayed stable merge.
5684 	 *
5685 	 * A delayed merge is scheduled (instead of performing an
5686 	 * early merge), in case bfqq might soon prove to be more
5687 	 * throughput-beneficial if not merged. Currently this is
5688 	 * possible only if bfqd is rotational with no queueing. For
5689 	 * such a drive, not merging bfqq is better for throughput if
5690 	 * bfqq happens to contain sequential I/O. So, we wait a
5691 	 * little bit for enough I/O to flow through bfqq. After that,
5692 	 * if such an I/O is sequential, then the merge is
5693 	 * canceled. Otherwise the merge is finally performed.
5694 	 */
5695 	if (!last_bfqq_created ||
5696 	    time_before(last_bfqq_created->creation_time +
5697 			msecs_to_jiffies(bfq_activation_stable_merging),
5698 			bfqq->creation_time) ||
5699 		bfqq->entity.parent != last_bfqq_created->entity.parent ||
5700 		bfqq->ioprio != last_bfqq_created->ioprio ||
5701 		bfqq->ioprio_class != last_bfqq_created->ioprio_class)
5702 		*source_bfqq = bfqq;
5703 	else if (time_after_eq(last_bfqq_created->creation_time +
5704 				 bfqd->bfq_burst_interval,
5705 				 bfqq->creation_time)) {
5706 		if (likely(bfqd->nonrot_with_queueing))
5707 			/*
5708 			 * With this type of drive, leaving
5709 			 * bfqq alone may provide no
5710 			 * throughput benefits compared with
5711 			 * merging bfqq. So merge bfqq now.
5712 			 */
5713 			bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5714 							 bic,
5715 							 last_bfqq_created);
5716 		else { /* schedule tentative stable merge */
5717 			/*
5718 			 * get reference on last_bfqq_created,
5719 			 * to prevent it from being freed,
5720 			 * until we decide whether to merge
5721 			 */
5722 			last_bfqq_created->ref++;
5723 			/*
5724 			 * need to keep track of stable refs, to
5725 			 * compute process refs correctly
5726 			 */
5727 			last_bfqq_created->stable_ref++;
5728 			/*
5729 			 * Record the bfqq to merge to.
5730 			 */
5731 			bic->stable_merge_bfqq = last_bfqq_created;
5732 		}
5733 	}
5734 
5735 	return bfqq;
5736 }
5737 
5738 
bfq_get_queue(struct bfq_data * bfqd,struct bio * bio,bool is_sync,struct bfq_io_cq * bic,bool respawn)5739 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5740 				       struct bio *bio, bool is_sync,
5741 				       struct bfq_io_cq *bic,
5742 				       bool respawn)
5743 {
5744 	const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5745 	const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5746 	struct bfq_queue **async_bfqq = NULL;
5747 	struct bfq_queue *bfqq;
5748 	struct bfq_group *bfqg;
5749 
5750 	bfqg = bfq_bio_bfqg(bfqd, bio);
5751 	if (!is_sync) {
5752 		async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5753 						  ioprio);
5754 		bfqq = *async_bfqq;
5755 		if (bfqq)
5756 			goto out;
5757 	}
5758 
5759 	bfqq = kmem_cache_alloc_node(bfq_pool,
5760 				     GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5761 				     bfqd->queue->node);
5762 
5763 	if (bfqq) {
5764 		bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5765 			      is_sync);
5766 		bfq_init_entity(&bfqq->entity, bfqg);
5767 		bfq_log_bfqq(bfqd, bfqq, "allocated");
5768 	} else {
5769 		bfqq = &bfqd->oom_bfqq;
5770 		bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5771 		goto out;
5772 	}
5773 
5774 	/*
5775 	 * Pin the queue now that it's allocated, scheduler exit will
5776 	 * prune it.
5777 	 */
5778 	if (async_bfqq) {
5779 		bfqq->ref++; /*
5780 			      * Extra group reference, w.r.t. sync
5781 			      * queue. This extra reference is removed
5782 			      * only if bfqq->bfqg disappears, to
5783 			      * guarantee that this queue is not freed
5784 			      * until its group goes away.
5785 			      */
5786 		bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5787 			     bfqq, bfqq->ref);
5788 		*async_bfqq = bfqq;
5789 	}
5790 
5791 out:
5792 	bfqq->ref++; /* get a process reference to this queue */
5793 
5794 	if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5795 		bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5796 	return bfqq;
5797 }
5798 
bfq_update_io_thinktime(struct bfq_data * bfqd,struct bfq_queue * bfqq)5799 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5800 				    struct bfq_queue *bfqq)
5801 {
5802 	struct bfq_ttime *ttime = &bfqq->ttime;
5803 	u64 elapsed;
5804 
5805 	/*
5806 	 * We are really interested in how long it takes for the queue to
5807 	 * become busy when there is no outstanding IO for this queue. So
5808 	 * ignore cases when the bfq queue has already IO queued.
5809 	 */
5810 	if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5811 		return;
5812 	elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5813 	elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5814 
5815 	ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5816 	ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
5817 	ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5818 				     ttime->ttime_samples);
5819 }
5820 
5821 static void
bfq_update_io_seektime(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct request * rq)5822 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5823 		       struct request *rq)
5824 {
5825 	bfqq->seek_history <<= 1;
5826 	bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5827 
5828 	if (bfqq->wr_coeff > 1 &&
5829 	    bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5830 	    BFQQ_TOTALLY_SEEKY(bfqq)) {
5831 		if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5832 					   bfq_wr_duration(bfqd))) {
5833 			/*
5834 			 * In soft_rt weight raising with the
5835 			 * interactive-weight-raising period
5836 			 * elapsed (so no switch back to
5837 			 * interactive weight raising).
5838 			 */
5839 			bfq_bfqq_end_wr(bfqq);
5840 		} else { /*
5841 			  * stopping soft_rt weight raising
5842 			  * while still in interactive period,
5843 			  * switch back to interactive weight
5844 			  * raising
5845 			  */
5846 			switch_back_to_interactive_wr(bfqq, bfqd);
5847 			bfqq->entity.prio_changed = 1;
5848 		}
5849 	}
5850 }
5851 
bfq_update_has_short_ttime(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct bfq_io_cq * bic)5852 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5853 				       struct bfq_queue *bfqq,
5854 				       struct bfq_io_cq *bic)
5855 {
5856 	bool has_short_ttime = true, state_changed;
5857 
5858 	/*
5859 	 * No need to update has_short_ttime if bfqq is async or in
5860 	 * idle io prio class, or if bfq_slice_idle is zero, because
5861 	 * no device idling is performed for bfqq in this case.
5862 	 */
5863 	if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5864 	    bfqd->bfq_slice_idle == 0)
5865 		return;
5866 
5867 	/* Idle window just restored, statistics are meaningless. */
5868 	if (time_is_after_eq_jiffies(bfqq->split_time +
5869 				     bfqd->bfq_wr_min_idle_time))
5870 		return;
5871 
5872 	/* Think time is infinite if no process is linked to
5873 	 * bfqq. Otherwise check average think time to decide whether
5874 	 * to mark as has_short_ttime. To this goal, compare average
5875 	 * think time with half the I/O-plugging timeout.
5876 	 */
5877 	if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5878 	    (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5879 	     bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5880 		has_short_ttime = false;
5881 
5882 	state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5883 
5884 	if (has_short_ttime)
5885 		bfq_mark_bfqq_has_short_ttime(bfqq);
5886 	else
5887 		bfq_clear_bfqq_has_short_ttime(bfqq);
5888 
5889 	/*
5890 	 * Until the base value for the total service time gets
5891 	 * finally computed for bfqq, the inject limit does depend on
5892 	 * the think-time state (short|long). In particular, the limit
5893 	 * is 0 or 1 if the think time is deemed, respectively, as
5894 	 * short or long (details in the comments in
5895 	 * bfq_update_inject_limit()). Accordingly, the next
5896 	 * instructions reset the inject limit if the think-time state
5897 	 * has changed and the above base value is still to be
5898 	 * computed.
5899 	 *
5900 	 * However, the reset is performed only if more than 100 ms
5901 	 * have elapsed since the last update of the inject limit, or
5902 	 * (inclusive) if the change is from short to long think
5903 	 * time. The reason for this waiting is as follows.
5904 	 *
5905 	 * bfqq may have a long think time because of a
5906 	 * synchronization with some other queue, i.e., because the
5907 	 * I/O of some other queue may need to be completed for bfqq
5908 	 * to receive new I/O. Details in the comments on the choice
5909 	 * of the queue for injection in bfq_select_queue().
5910 	 *
5911 	 * As stressed in those comments, if such a synchronization is
5912 	 * actually in place, then, without injection on bfqq, the
5913 	 * blocking I/O cannot happen to served while bfqq is in
5914 	 * service. As a consequence, if bfqq is granted
5915 	 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5916 	 * is dispatched, until the idle timeout fires. This is likely
5917 	 * to result in lower bandwidth and higher latencies for bfqq,
5918 	 * and in a severe loss of total throughput.
5919 	 *
5920 	 * On the opposite end, a non-zero inject limit may allow the
5921 	 * I/O that blocks bfqq to be executed soon, and therefore
5922 	 * bfqq to receive new I/O soon.
5923 	 *
5924 	 * But, if the blocking gets actually eliminated, then the
5925 	 * next think-time sample for bfqq may be very low. This in
5926 	 * turn may cause bfqq's think time to be deemed
5927 	 * short. Without the 100 ms barrier, this new state change
5928 	 * would cause the body of the next if to be executed
5929 	 * immediately. But this would set to 0 the inject
5930 	 * limit. Without injection, the blocking I/O would cause the
5931 	 * think time of bfqq to become long again, and therefore the
5932 	 * inject limit to be raised again, and so on. The only effect
5933 	 * of such a steady oscillation between the two think-time
5934 	 * states would be to prevent effective injection on bfqq.
5935 	 *
5936 	 * In contrast, if the inject limit is not reset during such a
5937 	 * long time interval as 100 ms, then the number of short
5938 	 * think time samples can grow significantly before the reset
5939 	 * is performed. As a consequence, the think time state can
5940 	 * become stable before the reset. Therefore there will be no
5941 	 * state change when the 100 ms elapse, and no reset of the
5942 	 * inject limit. The inject limit remains steadily equal to 1
5943 	 * both during and after the 100 ms. So injection can be
5944 	 * performed at all times, and throughput gets boosted.
5945 	 *
5946 	 * An inject limit equal to 1 is however in conflict, in
5947 	 * general, with the fact that the think time of bfqq is
5948 	 * short, because injection may be likely to delay bfqq's I/O
5949 	 * (as explained in the comments in
5950 	 * bfq_update_inject_limit()). But this does not happen in
5951 	 * this special case, because bfqq's low think time is due to
5952 	 * an effective handling of a synchronization, through
5953 	 * injection. In this special case, bfqq's I/O does not get
5954 	 * delayed by injection; on the contrary, bfqq's I/O is
5955 	 * brought forward, because it is not blocked for
5956 	 * milliseconds.
5957 	 *
5958 	 * In addition, serving the blocking I/O much sooner, and much
5959 	 * more frequently than once per I/O-plugging timeout, makes
5960 	 * it much quicker to detect a waker queue (the concept of
5961 	 * waker queue is defined in the comments in
5962 	 * bfq_add_request()). This makes it possible to start sooner
5963 	 * to boost throughput more effectively, by injecting the I/O
5964 	 * of the waker queue unconditionally on every
5965 	 * bfq_dispatch_request().
5966 	 *
5967 	 * One last, important benefit of not resetting the inject
5968 	 * limit before 100 ms is that, during this time interval, the
5969 	 * base value for the total service time is likely to get
5970 	 * finally computed for bfqq, freeing the inject limit from
5971 	 * its relation with the think time.
5972 	 */
5973 	if (state_changed && bfqq->last_serv_time_ns == 0 &&
5974 	    (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5975 				      msecs_to_jiffies(100)) ||
5976 	     !has_short_ttime))
5977 		bfq_reset_inject_limit(bfqd, bfqq);
5978 }
5979 
5980 /*
5981  * Called when a new fs request (rq) is added to bfqq.  Check if there's
5982  * something we should do about it.
5983  */
bfq_rq_enqueued(struct bfq_data * bfqd,struct bfq_queue * bfqq,struct request * rq)5984 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5985 			    struct request *rq)
5986 {
5987 	if (rq->cmd_flags & REQ_META)
5988 		bfqq->meta_pending++;
5989 
5990 	bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5991 
5992 	if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5993 		bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5994 				 blk_rq_sectors(rq) < 32;
5995 		bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5996 
5997 		/*
5998 		 * There is just this request queued: if
5999 		 * - the request is small, and
6000 		 * - we are idling to boost throughput, and
6001 		 * - the queue is not to be expired,
6002 		 * then just exit.
6003 		 *
6004 		 * In this way, if the device is being idled to wait
6005 		 * for a new request from the in-service queue, we
6006 		 * avoid unplugging the device and committing the
6007 		 * device to serve just a small request. In contrast
6008 		 * we wait for the block layer to decide when to
6009 		 * unplug the device: hopefully, new requests will be
6010 		 * merged to this one quickly, then the device will be
6011 		 * unplugged and larger requests will be dispatched.
6012 		 */
6013 		if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
6014 		    !budget_timeout)
6015 			return;
6016 
6017 		/*
6018 		 * A large enough request arrived, or idling is being
6019 		 * performed to preserve service guarantees, or
6020 		 * finally the queue is to be expired: in all these
6021 		 * cases disk idling is to be stopped, so clear
6022 		 * wait_request flag and reset timer.
6023 		 */
6024 		bfq_clear_bfqq_wait_request(bfqq);
6025 		hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
6026 
6027 		/*
6028 		 * The queue is not empty, because a new request just
6029 		 * arrived. Hence we can safely expire the queue, in
6030 		 * case of budget timeout, without risking that the
6031 		 * timestamps of the queue are not updated correctly.
6032 		 * See [1] for more details.
6033 		 */
6034 		if (budget_timeout)
6035 			bfq_bfqq_expire(bfqd, bfqq, false,
6036 					BFQQE_BUDGET_TIMEOUT);
6037 	}
6038 }
6039 
bfqq_request_allocated(struct bfq_queue * bfqq)6040 static void bfqq_request_allocated(struct bfq_queue *bfqq)
6041 {
6042 	struct bfq_entity *entity = &bfqq->entity;
6043 
6044 	for_each_entity(entity)
6045 		entity->allocated++;
6046 }
6047 
bfqq_request_freed(struct bfq_queue * bfqq)6048 static void bfqq_request_freed(struct bfq_queue *bfqq)
6049 {
6050 	struct bfq_entity *entity = &bfqq->entity;
6051 
6052 	for_each_entity(entity)
6053 		entity->allocated--;
6054 }
6055 
6056 /* returns true if it causes the idle timer to be disabled */
__bfq_insert_request(struct bfq_data * bfqd,struct request * rq)6057 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
6058 {
6059 	struct bfq_queue *bfqq = RQ_BFQQ(rq),
6060 		*new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
6061 						 RQ_BIC(rq));
6062 	bool waiting, idle_timer_disabled = false;
6063 
6064 	if (new_bfqq) {
6065 		/*
6066 		 * Release the request's reference to the old bfqq
6067 		 * and make sure one is taken to the shared queue.
6068 		 */
6069 		bfqq_request_allocated(new_bfqq);
6070 		bfqq_request_freed(bfqq);
6071 		new_bfqq->ref++;
6072 		/*
6073 		 * If the bic associated with the process
6074 		 * issuing this request still points to bfqq
6075 		 * (and thus has not been already redirected
6076 		 * to new_bfqq or even some other bfq_queue),
6077 		 * then complete the merge and redirect it to
6078 		 * new_bfqq.
6079 		 */
6080 		if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
6081 			bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
6082 					bfqq, new_bfqq);
6083 
6084 		bfq_clear_bfqq_just_created(bfqq);
6085 		/*
6086 		 * rq is about to be enqueued into new_bfqq,
6087 		 * release rq reference on bfqq
6088 		 */
6089 		bfq_put_queue(bfqq);
6090 		rq->elv.priv[1] = new_bfqq;
6091 		bfqq = new_bfqq;
6092 	}
6093 
6094 	bfq_update_io_thinktime(bfqd, bfqq);
6095 	bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6096 	bfq_update_io_seektime(bfqd, bfqq, rq);
6097 
6098 	waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6099 	bfq_add_request(rq);
6100 	idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6101 
6102 	rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6103 	list_add_tail(&rq->queuelist, &bfqq->fifo);
6104 
6105 	bfq_rq_enqueued(bfqd, bfqq, rq);
6106 
6107 	return idle_timer_disabled;
6108 }
6109 
6110 #ifdef CONFIG_BFQ_CGROUP_DEBUG
bfq_update_insert_stats(struct request_queue * q,struct bfq_queue * bfqq,bool idle_timer_disabled,blk_opf_t cmd_flags)6111 static void bfq_update_insert_stats(struct request_queue *q,
6112 				    struct bfq_queue *bfqq,
6113 				    bool idle_timer_disabled,
6114 				    blk_opf_t cmd_flags)
6115 {
6116 	if (!bfqq)
6117 		return;
6118 
6119 	/*
6120 	 * bfqq still exists, because it can disappear only after
6121 	 * either it is merged with another queue, or the process it
6122 	 * is associated with exits. But both actions must be taken by
6123 	 * the same process currently executing this flow of
6124 	 * instructions.
6125 	 *
6126 	 * In addition, the following queue lock guarantees that
6127 	 * bfqq_group(bfqq) exists as well.
6128 	 */
6129 	spin_lock_irq(&q->queue_lock);
6130 	bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6131 	if (idle_timer_disabled)
6132 		bfqg_stats_update_idle_time(bfqq_group(bfqq));
6133 	spin_unlock_irq(&q->queue_lock);
6134 }
6135 #else
bfq_update_insert_stats(struct request_queue * q,struct bfq_queue * bfqq,bool idle_timer_disabled,blk_opf_t cmd_flags)6136 static inline void bfq_update_insert_stats(struct request_queue *q,
6137 					   struct bfq_queue *bfqq,
6138 					   bool idle_timer_disabled,
6139 					   blk_opf_t cmd_flags) {}
6140 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
6141 
6142 static struct bfq_queue *bfq_init_rq(struct request *rq);
6143 
bfq_insert_request(struct blk_mq_hw_ctx * hctx,struct request * rq,bool at_head)6144 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6145 			       bool at_head)
6146 {
6147 	struct request_queue *q = hctx->queue;
6148 	struct bfq_data *bfqd = q->elevator->elevator_data;
6149 	struct bfq_queue *bfqq;
6150 	bool idle_timer_disabled = false;
6151 	blk_opf_t cmd_flags;
6152 	LIST_HEAD(free);
6153 
6154 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6155 	if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6156 		bfqg_stats_update_legacy_io(q, rq);
6157 #endif
6158 	spin_lock_irq(&bfqd->lock);
6159 	bfqq = bfq_init_rq(rq);
6160 	if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6161 		spin_unlock_irq(&bfqd->lock);
6162 		blk_mq_free_requests(&free);
6163 		return;
6164 	}
6165 
6166 	trace_block_rq_insert(rq);
6167 
6168 	if (!bfqq || at_head) {
6169 		if (at_head)
6170 			list_add(&rq->queuelist, &bfqd->dispatch);
6171 		else
6172 			list_add_tail(&rq->queuelist, &bfqd->dispatch);
6173 	} else {
6174 		idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6175 		/*
6176 		 * Update bfqq, because, if a queue merge has occurred
6177 		 * in __bfq_insert_request, then rq has been
6178 		 * redirected into a new queue.
6179 		 */
6180 		bfqq = RQ_BFQQ(rq);
6181 
6182 		if (rq_mergeable(rq)) {
6183 			elv_rqhash_add(q, rq);
6184 			if (!q->last_merge)
6185 				q->last_merge = rq;
6186 		}
6187 	}
6188 
6189 	/*
6190 	 * Cache cmd_flags before releasing scheduler lock, because rq
6191 	 * may disappear afterwards (for example, because of a request
6192 	 * merge).
6193 	 */
6194 	cmd_flags = rq->cmd_flags;
6195 	spin_unlock_irq(&bfqd->lock);
6196 
6197 	bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6198 				cmd_flags);
6199 }
6200 
bfq_insert_requests(struct blk_mq_hw_ctx * hctx,struct list_head * list,bool at_head)6201 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6202 				struct list_head *list, bool at_head)
6203 {
6204 	while (!list_empty(list)) {
6205 		struct request *rq;
6206 
6207 		rq = list_first_entry(list, struct request, queuelist);
6208 		list_del_init(&rq->queuelist);
6209 		bfq_insert_request(hctx, rq, at_head);
6210 	}
6211 }
6212 
bfq_update_hw_tag(struct bfq_data * bfqd)6213 static void bfq_update_hw_tag(struct bfq_data *bfqd)
6214 {
6215 	struct bfq_queue *bfqq = bfqd->in_service_queue;
6216 
6217 	bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6218 				       bfqd->rq_in_driver);
6219 
6220 	if (bfqd->hw_tag == 1)
6221 		return;
6222 
6223 	/*
6224 	 * This sample is valid if the number of outstanding requests
6225 	 * is large enough to allow a queueing behavior.  Note that the
6226 	 * sum is not exact, as it's not taking into account deactivated
6227 	 * requests.
6228 	 */
6229 	if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6230 		return;
6231 
6232 	/*
6233 	 * If active queue hasn't enough requests and can idle, bfq might not
6234 	 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6235 	 * case
6236 	 */
6237 	if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6238 	    bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6239 	    BFQ_HW_QUEUE_THRESHOLD &&
6240 	    bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6241 		return;
6242 
6243 	if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6244 		return;
6245 
6246 	bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6247 	bfqd->max_rq_in_driver = 0;
6248 	bfqd->hw_tag_samples = 0;
6249 
6250 	bfqd->nonrot_with_queueing =
6251 		blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6252 }
6253 
bfq_completed_request(struct bfq_queue * bfqq,struct bfq_data * bfqd)6254 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6255 {
6256 	u64 now_ns;
6257 	u32 delta_us;
6258 
6259 	bfq_update_hw_tag(bfqd);
6260 
6261 	bfqd->rq_in_driver--;
6262 	bfqq->dispatched--;
6263 
6264 	if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6265 		/*
6266 		 * Set budget_timeout (which we overload to store the
6267 		 * time at which the queue remains with no backlog and
6268 		 * no outstanding request; used by the weight-raising
6269 		 * mechanism).
6270 		 */
6271 		bfqq->budget_timeout = jiffies;
6272 
6273 		bfq_weights_tree_remove(bfqd, bfqq);
6274 	}
6275 
6276 	now_ns = ktime_get_ns();
6277 
6278 	bfqq->ttime.last_end_request = now_ns;
6279 
6280 	/*
6281 	 * Using us instead of ns, to get a reasonable precision in
6282 	 * computing rate in next check.
6283 	 */
6284 	delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6285 
6286 	/*
6287 	 * If the request took rather long to complete, and, according
6288 	 * to the maximum request size recorded, this completion latency
6289 	 * implies that the request was certainly served at a very low
6290 	 * rate (less than 1M sectors/sec), then the whole observation
6291 	 * interval that lasts up to this time instant cannot be a
6292 	 * valid time interval for computing a new peak rate.  Invoke
6293 	 * bfq_update_rate_reset to have the following three steps
6294 	 * taken:
6295 	 * - close the observation interval at the last (previous)
6296 	 *   request dispatch or completion
6297 	 * - compute rate, if possible, for that observation interval
6298 	 * - reset to zero samples, which will trigger a proper
6299 	 *   re-initialization of the observation interval on next
6300 	 *   dispatch
6301 	 */
6302 	if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6303 	   (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6304 			1UL<<(BFQ_RATE_SHIFT - 10))
6305 		bfq_update_rate_reset(bfqd, NULL);
6306 	bfqd->last_completion = now_ns;
6307 	/*
6308 	 * Shared queues are likely to receive I/O at a high
6309 	 * rate. This may deceptively let them be considered as wakers
6310 	 * of other queues. But a false waker will unjustly steal
6311 	 * bandwidth to its supposedly woken queue. So considering
6312 	 * also shared queues in the waking mechanism may cause more
6313 	 * control troubles than throughput benefits. Then reset
6314 	 * last_completed_rq_bfqq if bfqq is a shared queue.
6315 	 */
6316 	if (!bfq_bfqq_coop(bfqq))
6317 		bfqd->last_completed_rq_bfqq = bfqq;
6318 	else
6319 		bfqd->last_completed_rq_bfqq = NULL;
6320 
6321 	/*
6322 	 * If we are waiting to discover whether the request pattern
6323 	 * of the task associated with the queue is actually
6324 	 * isochronous, and both requisites for this condition to hold
6325 	 * are now satisfied, then compute soft_rt_next_start (see the
6326 	 * comments on the function bfq_bfqq_softrt_next_start()). We
6327 	 * do not compute soft_rt_next_start if bfqq is in interactive
6328 	 * weight raising (see the comments in bfq_bfqq_expire() for
6329 	 * an explanation). We schedule this delayed update when bfqq
6330 	 * expires, if it still has in-flight requests.
6331 	 */
6332 	if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6333 	    RB_EMPTY_ROOT(&bfqq->sort_list) &&
6334 	    bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6335 		bfqq->soft_rt_next_start =
6336 			bfq_bfqq_softrt_next_start(bfqd, bfqq);
6337 
6338 	/*
6339 	 * If this is the in-service queue, check if it needs to be expired,
6340 	 * or if we want to idle in case it has no pending requests.
6341 	 */
6342 	if (bfqd->in_service_queue == bfqq) {
6343 		if (bfq_bfqq_must_idle(bfqq)) {
6344 			if (bfqq->dispatched == 0)
6345 				bfq_arm_slice_timer(bfqd);
6346 			/*
6347 			 * If we get here, we do not expire bfqq, even
6348 			 * if bfqq was in budget timeout or had no
6349 			 * more requests (as controlled in the next
6350 			 * conditional instructions). The reason for
6351 			 * not expiring bfqq is as follows.
6352 			 *
6353 			 * Here bfqq->dispatched > 0 holds, but
6354 			 * bfq_bfqq_must_idle() returned true. This
6355 			 * implies that, even if no request arrives
6356 			 * for bfqq before bfqq->dispatched reaches 0,
6357 			 * bfqq will, however, not be expired on the
6358 			 * completion event that causes bfqq->dispatch
6359 			 * to reach zero. In contrast, on this event,
6360 			 * bfqq will start enjoying device idling
6361 			 * (I/O-dispatch plugging).
6362 			 *
6363 			 * But, if we expired bfqq here, bfqq would
6364 			 * not have the chance to enjoy device idling
6365 			 * when bfqq->dispatched finally reaches
6366 			 * zero. This would expose bfqq to violation
6367 			 * of its reserved service guarantees.
6368 			 */
6369 			return;
6370 		} else if (bfq_may_expire_for_budg_timeout(bfqq))
6371 			bfq_bfqq_expire(bfqd, bfqq, false,
6372 					BFQQE_BUDGET_TIMEOUT);
6373 		else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6374 			 (bfqq->dispatched == 0 ||
6375 			  !bfq_better_to_idle(bfqq)))
6376 			bfq_bfqq_expire(bfqd, bfqq, false,
6377 					BFQQE_NO_MORE_REQUESTS);
6378 	}
6379 
6380 	if (!bfqd->rq_in_driver)
6381 		bfq_schedule_dispatch(bfqd);
6382 }
6383 
6384 /*
6385  * The processes associated with bfqq may happen to generate their
6386  * cumulative I/O at a lower rate than the rate at which the device
6387  * could serve the same I/O. This is rather probable, e.g., if only
6388  * one process is associated with bfqq and the device is an SSD. It
6389  * results in bfqq becoming often empty while in service. In this
6390  * respect, if BFQ is allowed to switch to another queue when bfqq
6391  * remains empty, then the device goes on being fed with I/O requests,
6392  * and the throughput is not affected. In contrast, if BFQ is not
6393  * allowed to switch to another queue---because bfqq is sync and
6394  * I/O-dispatch needs to be plugged while bfqq is temporarily
6395  * empty---then, during the service of bfqq, there will be frequent
6396  * "service holes", i.e., time intervals during which bfqq gets empty
6397  * and the device can only consume the I/O already queued in its
6398  * hardware queues. During service holes, the device may even get to
6399  * remaining idle. In the end, during the service of bfqq, the device
6400  * is driven at a lower speed than the one it can reach with the kind
6401  * of I/O flowing through bfqq.
6402  *
6403  * To counter this loss of throughput, BFQ implements a "request
6404  * injection mechanism", which tries to fill the above service holes
6405  * with I/O requests taken from other queues. The hard part in this
6406  * mechanism is finding the right amount of I/O to inject, so as to
6407  * both boost throughput and not break bfqq's bandwidth and latency
6408  * guarantees. In this respect, the mechanism maintains a per-queue
6409  * inject limit, computed as below. While bfqq is empty, the injection
6410  * mechanism dispatches extra I/O requests only until the total number
6411  * of I/O requests in flight---i.e., already dispatched but not yet
6412  * completed---remains lower than this limit.
6413  *
6414  * A first definition comes in handy to introduce the algorithm by
6415  * which the inject limit is computed.  We define as first request for
6416  * bfqq, an I/O request for bfqq that arrives while bfqq is in
6417  * service, and causes bfqq to switch from empty to non-empty. The
6418  * algorithm updates the limit as a function of the effect of
6419  * injection on the service times of only the first requests of
6420  * bfqq. The reason for this restriction is that these are the
6421  * requests whose service time is affected most, because they are the
6422  * first to arrive after injection possibly occurred.
6423  *
6424  * To evaluate the effect of injection, the algorithm measures the
6425  * "total service time" of first requests. We define as total service
6426  * time of an I/O request, the time that elapses since when the
6427  * request is enqueued into bfqq, to when it is completed. This
6428  * quantity allows the whole effect of injection to be measured. It is
6429  * easy to see why. Suppose that some requests of other queues are
6430  * actually injected while bfqq is empty, and that a new request R
6431  * then arrives for bfqq. If the device does start to serve all or
6432  * part of the injected requests during the service hole, then,
6433  * because of this extra service, it may delay the next invocation of
6434  * the dispatch hook of BFQ. Then, even after R gets eventually
6435  * dispatched, the device may delay the actual service of R if it is
6436  * still busy serving the extra requests, or if it decides to serve,
6437  * before R, some extra request still present in its queues. As a
6438  * conclusion, the cumulative extra delay caused by injection can be
6439  * easily evaluated by just comparing the total service time of first
6440  * requests with and without injection.
6441  *
6442  * The limit-update algorithm works as follows. On the arrival of a
6443  * first request of bfqq, the algorithm measures the total time of the
6444  * request only if one of the three cases below holds, and, for each
6445  * case, it updates the limit as described below:
6446  *
6447  * (1) If there is no in-flight request. This gives a baseline for the
6448  *     total service time of the requests of bfqq. If the baseline has
6449  *     not been computed yet, then, after computing it, the limit is
6450  *     set to 1, to start boosting throughput, and to prepare the
6451  *     ground for the next case. If the baseline has already been
6452  *     computed, then it is updated, in case it results to be lower
6453  *     than the previous value.
6454  *
6455  * (2) If the limit is higher than 0 and there are in-flight
6456  *     requests. By comparing the total service time in this case with
6457  *     the above baseline, it is possible to know at which extent the
6458  *     current value of the limit is inflating the total service
6459  *     time. If the inflation is below a certain threshold, then bfqq
6460  *     is assumed to be suffering from no perceivable loss of its
6461  *     service guarantees, and the limit is even tentatively
6462  *     increased. If the inflation is above the threshold, then the
6463  *     limit is decreased. Due to the lack of any hysteresis, this
6464  *     logic makes the limit oscillate even in steady workload
6465  *     conditions. Yet we opted for it, because it is fast in reaching
6466  *     the best value for the limit, as a function of the current I/O
6467  *     workload. To reduce oscillations, this step is disabled for a
6468  *     short time interval after the limit happens to be decreased.
6469  *
6470  * (3) Periodically, after resetting the limit, to make sure that the
6471  *     limit eventually drops in case the workload changes. This is
6472  *     needed because, after the limit has gone safely up for a
6473  *     certain workload, it is impossible to guess whether the
6474  *     baseline total service time may have changed, without measuring
6475  *     it again without injection. A more effective version of this
6476  *     step might be to just sample the baseline, by interrupting
6477  *     injection only once, and then to reset/lower the limit only if
6478  *     the total service time with the current limit does happen to be
6479  *     too large.
6480  *
6481  * More details on each step are provided in the comments on the
6482  * pieces of code that implement these steps: the branch handling the
6483  * transition from empty to non empty in bfq_add_request(), the branch
6484  * handling injection in bfq_select_queue(), and the function
6485  * bfq_choose_bfqq_for_injection(). These comments also explain some
6486  * exceptions, made by the injection mechanism in some special cases.
6487  */
bfq_update_inject_limit(struct bfq_data * bfqd,struct bfq_queue * bfqq)6488 static void bfq_update_inject_limit(struct bfq_data *bfqd,
6489 				    struct bfq_queue *bfqq)
6490 {
6491 	u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
6492 	unsigned int old_limit = bfqq->inject_limit;
6493 
6494 	if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6495 		u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6496 
6497 		if (tot_time_ns >= threshold && old_limit > 0) {
6498 			bfqq->inject_limit--;
6499 			bfqq->decrease_time_jif = jiffies;
6500 		} else if (tot_time_ns < threshold &&
6501 			   old_limit <= bfqd->max_rq_in_driver)
6502 			bfqq->inject_limit++;
6503 	}
6504 
6505 	/*
6506 	 * Either we still have to compute the base value for the
6507 	 * total service time, and there seem to be the right
6508 	 * conditions to do it, or we can lower the last base value
6509 	 * computed.
6510 	 *
6511 	 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6512 	 * request in flight, because this function is in the code
6513 	 * path that handles the completion of a request of bfqq, and,
6514 	 * in particular, this function is executed before
6515 	 * bfqd->rq_in_driver is decremented in such a code path.
6516 	 */
6517 	if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
6518 	    tot_time_ns < bfqq->last_serv_time_ns) {
6519 		if (bfqq->last_serv_time_ns == 0) {
6520 			/*
6521 			 * Now we certainly have a base value: make sure we
6522 			 * start trying injection.
6523 			 */
6524 			bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6525 		}
6526 		bfqq->last_serv_time_ns = tot_time_ns;
6527 	} else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
6528 		/*
6529 		 * No I/O injected and no request still in service in
6530 		 * the drive: these are the exact conditions for
6531 		 * computing the base value of the total service time
6532 		 * for bfqq. So let's update this value, because it is
6533 		 * rather variable. For example, it varies if the size
6534 		 * or the spatial locality of the I/O requests in bfqq
6535 		 * change.
6536 		 */
6537 		bfqq->last_serv_time_ns = tot_time_ns;
6538 
6539 
6540 	/* update complete, not waiting for any request completion any longer */
6541 	bfqd->waited_rq = NULL;
6542 	bfqd->rqs_injected = false;
6543 }
6544 
6545 /*
6546  * Handle either a requeue or a finish for rq. The things to do are
6547  * the same in both cases: all references to rq are to be dropped. In
6548  * particular, rq is considered completed from the point of view of
6549  * the scheduler.
6550  */
bfq_finish_requeue_request(struct request * rq)6551 static void bfq_finish_requeue_request(struct request *rq)
6552 {
6553 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
6554 	struct bfq_data *bfqd;
6555 	unsigned long flags;
6556 
6557 	/*
6558 	 * rq either is not associated with any icq, or is an already
6559 	 * requeued request that has not (yet) been re-inserted into
6560 	 * a bfq_queue.
6561 	 */
6562 	if (!rq->elv.icq || !bfqq)
6563 		return;
6564 
6565 	bfqd = bfqq->bfqd;
6566 
6567 	if (rq->rq_flags & RQF_STARTED)
6568 		bfqg_stats_update_completion(bfqq_group(bfqq),
6569 					     rq->start_time_ns,
6570 					     rq->io_start_time_ns,
6571 					     rq->cmd_flags);
6572 
6573 	spin_lock_irqsave(&bfqd->lock, flags);
6574 	if (likely(rq->rq_flags & RQF_STARTED)) {
6575 		if (rq == bfqd->waited_rq)
6576 			bfq_update_inject_limit(bfqd, bfqq);
6577 
6578 		bfq_completed_request(bfqq, bfqd);
6579 	}
6580 	bfqq_request_freed(bfqq);
6581 	bfq_put_queue(bfqq);
6582 	RQ_BIC(rq)->requests--;
6583 	spin_unlock_irqrestore(&bfqd->lock, flags);
6584 
6585 	/*
6586 	 * Reset private fields. In case of a requeue, this allows
6587 	 * this function to correctly do nothing if it is spuriously
6588 	 * invoked again on this same request (see the check at the
6589 	 * beginning of the function). Probably, a better general
6590 	 * design would be to prevent blk-mq from invoking the requeue
6591 	 * or finish hooks of an elevator, for a request that is not
6592 	 * referred by that elevator.
6593 	 *
6594 	 * Resetting the following fields would break the
6595 	 * request-insertion logic if rq is re-inserted into a bfq
6596 	 * internal queue, without a re-preparation. Here we assume
6597 	 * that re-insertions of requeued requests, without
6598 	 * re-preparation, can happen only for pass_through or at_head
6599 	 * requests (which are not re-inserted into bfq internal
6600 	 * queues).
6601 	 */
6602 	rq->elv.priv[0] = NULL;
6603 	rq->elv.priv[1] = NULL;
6604 }
6605 
bfq_finish_request(struct request * rq)6606 static void bfq_finish_request(struct request *rq)
6607 {
6608 	bfq_finish_requeue_request(rq);
6609 
6610 	if (rq->elv.icq) {
6611 		put_io_context(rq->elv.icq->ioc);
6612 		rq->elv.icq = NULL;
6613 	}
6614 }
6615 
6616 /*
6617  * Removes the association between the current task and bfqq, assuming
6618  * that bic points to the bfq iocontext of the task.
6619  * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6620  * was the last process referring to that bfqq.
6621  */
6622 static struct bfq_queue *
bfq_split_bfqq(struct bfq_io_cq * bic,struct bfq_queue * bfqq)6623 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6624 {
6625 	bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6626 
6627 	if (bfqq_process_refs(bfqq) == 1) {
6628 		bfqq->pid = current->pid;
6629 		bfq_clear_bfqq_coop(bfqq);
6630 		bfq_clear_bfqq_split_coop(bfqq);
6631 		return bfqq;
6632 	}
6633 
6634 	bic_set_bfqq(bic, NULL, true);
6635 
6636 	bfq_put_cooperator(bfqq);
6637 
6638 	bfq_release_process_ref(bfqq->bfqd, bfqq);
6639 	return NULL;
6640 }
6641 
bfq_get_bfqq_handle_split(struct bfq_data * bfqd,struct bfq_io_cq * bic,struct bio * bio,bool split,bool is_sync,bool * new_queue)6642 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6643 						   struct bfq_io_cq *bic,
6644 						   struct bio *bio,
6645 						   bool split, bool is_sync,
6646 						   bool *new_queue)
6647 {
6648 	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
6649 
6650 	if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6651 		return bfqq;
6652 
6653 	if (new_queue)
6654 		*new_queue = true;
6655 
6656 	if (bfqq)
6657 		bfq_put_queue(bfqq);
6658 	bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6659 
6660 	bic_set_bfqq(bic, bfqq, is_sync);
6661 	if (split && is_sync) {
6662 		if ((bic->was_in_burst_list && bfqd->large_burst) ||
6663 		    bic->saved_in_large_burst)
6664 			bfq_mark_bfqq_in_large_burst(bfqq);
6665 		else {
6666 			bfq_clear_bfqq_in_large_burst(bfqq);
6667 			if (bic->was_in_burst_list)
6668 				/*
6669 				 * If bfqq was in the current
6670 				 * burst list before being
6671 				 * merged, then we have to add
6672 				 * it back. And we do not need
6673 				 * to increase burst_size, as
6674 				 * we did not decrement
6675 				 * burst_size when we removed
6676 				 * bfqq from the burst list as
6677 				 * a consequence of a merge
6678 				 * (see comments in
6679 				 * bfq_put_queue). In this
6680 				 * respect, it would be rather
6681 				 * costly to know whether the
6682 				 * current burst list is still
6683 				 * the same burst list from
6684 				 * which bfqq was removed on
6685 				 * the merge. To avoid this
6686 				 * cost, if bfqq was in a
6687 				 * burst list, then we add
6688 				 * bfqq to the current burst
6689 				 * list without any further
6690 				 * check. This can cause
6691 				 * inappropriate insertions,
6692 				 * but rarely enough to not
6693 				 * harm the detection of large
6694 				 * bursts significantly.
6695 				 */
6696 				hlist_add_head(&bfqq->burst_list_node,
6697 					       &bfqd->burst_list);
6698 		}
6699 		bfqq->split_time = jiffies;
6700 	}
6701 
6702 	return bfqq;
6703 }
6704 
6705 /*
6706  * Only reset private fields. The actual request preparation will be
6707  * performed by bfq_init_rq, when rq is either inserted or merged. See
6708  * comments on bfq_init_rq for the reason behind this delayed
6709  * preparation.
6710  */
bfq_prepare_request(struct request * rq)6711 static void bfq_prepare_request(struct request *rq)
6712 {
6713 	rq->elv.icq = ioc_find_get_icq(rq->q);
6714 
6715 	/*
6716 	 * Regardless of whether we have an icq attached, we have to
6717 	 * clear the scheduler pointers, as they might point to
6718 	 * previously allocated bic/bfqq structs.
6719 	 */
6720 	rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6721 }
6722 
6723 /*
6724  * If needed, init rq, allocate bfq data structures associated with
6725  * rq, and increment reference counters in the destination bfq_queue
6726  * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6727  * not associated with any bfq_queue.
6728  *
6729  * This function is invoked by the functions that perform rq insertion
6730  * or merging. One may have expected the above preparation operations
6731  * to be performed in bfq_prepare_request, and not delayed to when rq
6732  * is inserted or merged. The rationale behind this delayed
6733  * preparation is that, after the prepare_request hook is invoked for
6734  * rq, rq may still be transformed into a request with no icq, i.e., a
6735  * request not associated with any queue. No bfq hook is invoked to
6736  * signal this transformation. As a consequence, should these
6737  * preparation operations be performed when the prepare_request hook
6738  * is invoked, and should rq be transformed one moment later, bfq
6739  * would end up in an inconsistent state, because it would have
6740  * incremented some queue counters for an rq destined to
6741  * transformation, without any chance to correctly lower these
6742  * counters back. In contrast, no transformation can still happen for
6743  * rq after rq has been inserted or merged. So, it is safe to execute
6744  * these preparation operations when rq is finally inserted or merged.
6745  */
bfq_init_rq(struct request * rq)6746 static struct bfq_queue *bfq_init_rq(struct request *rq)
6747 {
6748 	struct request_queue *q = rq->q;
6749 	struct bio *bio = rq->bio;
6750 	struct bfq_data *bfqd = q->elevator->elevator_data;
6751 	struct bfq_io_cq *bic;
6752 	const int is_sync = rq_is_sync(rq);
6753 	struct bfq_queue *bfqq;
6754 	bool new_queue = false;
6755 	bool bfqq_already_existing = false, split = false;
6756 
6757 	if (unlikely(!rq->elv.icq))
6758 		return NULL;
6759 
6760 	/*
6761 	 * Assuming that elv.priv[1] is set only if everything is set
6762 	 * for this rq. This holds true, because this function is
6763 	 * invoked only for insertion or merging, and, after such
6764 	 * events, a request cannot be manipulated any longer before
6765 	 * being removed from bfq.
6766 	 */
6767 	if (rq->elv.priv[1])
6768 		return rq->elv.priv[1];
6769 
6770 	bic = icq_to_bic(rq->elv.icq);
6771 
6772 	bfq_check_ioprio_change(bic, bio);
6773 
6774 	bfq_bic_update_cgroup(bic, bio);
6775 
6776 	bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6777 					 &new_queue);
6778 
6779 	if (likely(!new_queue)) {
6780 		/* If the queue was seeky for too long, break it apart. */
6781 		if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
6782 			!bic->stably_merged) {
6783 			struct bfq_queue *old_bfqq = bfqq;
6784 
6785 			/* Update bic before losing reference to bfqq */
6786 			if (bfq_bfqq_in_large_burst(bfqq))
6787 				bic->saved_in_large_burst = true;
6788 
6789 			bfqq = bfq_split_bfqq(bic, bfqq);
6790 			split = true;
6791 
6792 			if (!bfqq) {
6793 				bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6794 								 true, is_sync,
6795 								 NULL);
6796 				if (unlikely(bfqq == &bfqd->oom_bfqq))
6797 					bfqq_already_existing = true;
6798 			} else
6799 				bfqq_already_existing = true;
6800 
6801 			if (!bfqq_already_existing) {
6802 				bfqq->waker_bfqq = old_bfqq->waker_bfqq;
6803 				bfqq->tentative_waker_bfqq = NULL;
6804 
6805 				/*
6806 				 * If the waker queue disappears, then
6807 				 * new_bfqq->waker_bfqq must be
6808 				 * reset. So insert new_bfqq into the
6809 				 * woken_list of the waker. See
6810 				 * bfq_check_waker for details.
6811 				 */
6812 				if (bfqq->waker_bfqq)
6813 					hlist_add_head(&bfqq->woken_list_node,
6814 						       &bfqq->waker_bfqq->woken_list);
6815 			}
6816 		}
6817 	}
6818 
6819 	bfqq_request_allocated(bfqq);
6820 	bfqq->ref++;
6821 	bic->requests++;
6822 	bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6823 		     rq, bfqq, bfqq->ref);
6824 
6825 	rq->elv.priv[0] = bic;
6826 	rq->elv.priv[1] = bfqq;
6827 
6828 	/*
6829 	 * If a bfq_queue has only one process reference, it is owned
6830 	 * by only this bic: we can then set bfqq->bic = bic. in
6831 	 * addition, if the queue has also just been split, we have to
6832 	 * resume its state.
6833 	 */
6834 	if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6835 		bfqq->bic = bic;
6836 		if (split) {
6837 			/*
6838 			 * The queue has just been split from a shared
6839 			 * queue: restore the idle window and the
6840 			 * possible weight raising period.
6841 			 */
6842 			bfq_bfqq_resume_state(bfqq, bfqd, bic,
6843 					      bfqq_already_existing);
6844 		}
6845 	}
6846 
6847 	/*
6848 	 * Consider bfqq as possibly belonging to a burst of newly
6849 	 * created queues only if:
6850 	 * 1) A burst is actually happening (bfqd->burst_size > 0)
6851 	 * or
6852 	 * 2) There is no other active queue. In fact, if, in
6853 	 *    contrast, there are active queues not belonging to the
6854 	 *    possible burst bfqq may belong to, then there is no gain
6855 	 *    in considering bfqq as belonging to a burst, and
6856 	 *    therefore in not weight-raising bfqq. See comments on
6857 	 *    bfq_handle_burst().
6858 	 *
6859 	 * This filtering also helps eliminating false positives,
6860 	 * occurring when bfqq does not belong to an actual large
6861 	 * burst, but some background task (e.g., a service) happens
6862 	 * to trigger the creation of new queues very close to when
6863 	 * bfqq and its possible companion queues are created. See
6864 	 * comments on bfq_handle_burst() for further details also on
6865 	 * this issue.
6866 	 */
6867 	if (unlikely(bfq_bfqq_just_created(bfqq) &&
6868 		     (bfqd->burst_size > 0 ||
6869 		      bfq_tot_busy_queues(bfqd) == 0)))
6870 		bfq_handle_burst(bfqd, bfqq);
6871 
6872 	return bfqq;
6873 }
6874 
6875 static void
bfq_idle_slice_timer_body(struct bfq_data * bfqd,struct bfq_queue * bfqq)6876 bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6877 {
6878 	enum bfqq_expiration reason;
6879 	unsigned long flags;
6880 
6881 	spin_lock_irqsave(&bfqd->lock, flags);
6882 
6883 	/*
6884 	 * Considering that bfqq may be in race, we should firstly check
6885 	 * whether bfqq is in service before doing something on it. If
6886 	 * the bfqq in race is not in service, it has already been expired
6887 	 * through __bfq_bfqq_expire func and its wait_request flags has
6888 	 * been cleared in __bfq_bfqd_reset_in_service func.
6889 	 */
6890 	if (bfqq != bfqd->in_service_queue) {
6891 		spin_unlock_irqrestore(&bfqd->lock, flags);
6892 		return;
6893 	}
6894 
6895 	bfq_clear_bfqq_wait_request(bfqq);
6896 
6897 	if (bfq_bfqq_budget_timeout(bfqq))
6898 		/*
6899 		 * Also here the queue can be safely expired
6900 		 * for budget timeout without wasting
6901 		 * guarantees
6902 		 */
6903 		reason = BFQQE_BUDGET_TIMEOUT;
6904 	else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6905 		/*
6906 		 * The queue may not be empty upon timer expiration,
6907 		 * because we may not disable the timer when the
6908 		 * first request of the in-service queue arrives
6909 		 * during disk idling.
6910 		 */
6911 		reason = BFQQE_TOO_IDLE;
6912 	else
6913 		goto schedule_dispatch;
6914 
6915 	bfq_bfqq_expire(bfqd, bfqq, true, reason);
6916 
6917 schedule_dispatch:
6918 	bfq_schedule_dispatch(bfqd);
6919 	spin_unlock_irqrestore(&bfqd->lock, flags);
6920 }
6921 
6922 /*
6923  * Handler of the expiration of the timer running if the in-service queue
6924  * is idling inside its time slice.
6925  */
bfq_idle_slice_timer(struct hrtimer * timer)6926 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6927 {
6928 	struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6929 					     idle_slice_timer);
6930 	struct bfq_queue *bfqq = bfqd->in_service_queue;
6931 
6932 	/*
6933 	 * Theoretical race here: the in-service queue can be NULL or
6934 	 * different from the queue that was idling if a new request
6935 	 * arrives for the current queue and there is a full dispatch
6936 	 * cycle that changes the in-service queue.  This can hardly
6937 	 * happen, but in the worst case we just expire a queue too
6938 	 * early.
6939 	 */
6940 	if (bfqq)
6941 		bfq_idle_slice_timer_body(bfqd, bfqq);
6942 
6943 	return HRTIMER_NORESTART;
6944 }
6945 
__bfq_put_async_bfqq(struct bfq_data * bfqd,struct bfq_queue ** bfqq_ptr)6946 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6947 				 struct bfq_queue **bfqq_ptr)
6948 {
6949 	struct bfq_queue *bfqq = *bfqq_ptr;
6950 
6951 	bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6952 	if (bfqq) {
6953 		bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6954 
6955 		bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6956 			     bfqq, bfqq->ref);
6957 		bfq_put_queue(bfqq);
6958 		*bfqq_ptr = NULL;
6959 	}
6960 }
6961 
6962 /*
6963  * Release all the bfqg references to its async queues.  If we are
6964  * deallocating the group these queues may still contain requests, so
6965  * we reparent them to the root cgroup (i.e., the only one that will
6966  * exist for sure until all the requests on a device are gone).
6967  */
bfq_put_async_queues(struct bfq_data * bfqd,struct bfq_group * bfqg)6968 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6969 {
6970 	int i, j;
6971 
6972 	for (i = 0; i < 2; i++)
6973 		for (j = 0; j < IOPRIO_NR_LEVELS; j++)
6974 			__bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6975 
6976 	__bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6977 }
6978 
6979 /*
6980  * See the comments on bfq_limit_depth for the purpose of
6981  * the depths set in the function. Return minimum shallow depth we'll use.
6982  */
bfq_update_depths(struct bfq_data * bfqd,struct sbitmap_queue * bt)6983 static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
6984 {
6985 	unsigned int depth = 1U << bt->sb.shift;
6986 
6987 	bfqd->full_depth_shift = bt->sb.shift;
6988 	/*
6989 	 * In-word depths if no bfq_queue is being weight-raised:
6990 	 * leaving 25% of tags only for sync reads.
6991 	 *
6992 	 * In next formulas, right-shift the value
6993 	 * (1U<<bt->sb.shift), instead of computing directly
6994 	 * (1U<<(bt->sb.shift - something)), to be robust against
6995 	 * any possible value of bt->sb.shift, without having to
6996 	 * limit 'something'.
6997 	 */
6998 	/* no more than 50% of tags for async I/O */
6999 	bfqd->word_depths[0][0] = max(depth >> 1, 1U);
7000 	/*
7001 	 * no more than 75% of tags for sync writes (25% extra tags
7002 	 * w.r.t. async I/O, to prevent async I/O from starving sync
7003 	 * writes)
7004 	 */
7005 	bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
7006 
7007 	/*
7008 	 * In-word depths in case some bfq_queue is being weight-
7009 	 * raised: leaving ~63% of tags for sync reads. This is the
7010 	 * highest percentage for which, in our tests, application
7011 	 * start-up times didn't suffer from any regression due to tag
7012 	 * shortage.
7013 	 */
7014 	/* no more than ~18% of tags for async I/O */
7015 	bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
7016 	/* no more than ~37% of tags for sync writes (~20% extra tags) */
7017 	bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
7018 }
7019 
bfq_depth_updated(struct blk_mq_hw_ctx * hctx)7020 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
7021 {
7022 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
7023 	struct blk_mq_tags *tags = hctx->sched_tags;
7024 
7025 	bfq_update_depths(bfqd, &tags->bitmap_tags);
7026 	sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
7027 }
7028 
bfq_init_hctx(struct blk_mq_hw_ctx * hctx,unsigned int index)7029 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
7030 {
7031 	bfq_depth_updated(hctx);
7032 	return 0;
7033 }
7034 
bfq_exit_queue(struct elevator_queue * e)7035 static void bfq_exit_queue(struct elevator_queue *e)
7036 {
7037 	struct bfq_data *bfqd = e->elevator_data;
7038 	struct bfq_queue *bfqq, *n;
7039 
7040 	hrtimer_cancel(&bfqd->idle_slice_timer);
7041 
7042 	spin_lock_irq(&bfqd->lock);
7043 	list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
7044 		bfq_deactivate_bfqq(bfqd, bfqq, false, false);
7045 	spin_unlock_irq(&bfqd->lock);
7046 
7047 	hrtimer_cancel(&bfqd->idle_slice_timer);
7048 
7049 	/* release oom-queue reference to root group */
7050 	bfqg_and_blkg_put(bfqd->root_group);
7051 
7052 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7053 	blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
7054 #else
7055 	spin_lock_irq(&bfqd->lock);
7056 	bfq_put_async_queues(bfqd, bfqd->root_group);
7057 	kfree(bfqd->root_group);
7058 	spin_unlock_irq(&bfqd->lock);
7059 #endif
7060 
7061 	blk_stat_disable_accounting(bfqd->queue);
7062 	wbt_enable_default(bfqd->queue);
7063 
7064 	kfree(bfqd);
7065 }
7066 
bfq_init_root_group(struct bfq_group * root_group,struct bfq_data * bfqd)7067 static void bfq_init_root_group(struct bfq_group *root_group,
7068 				struct bfq_data *bfqd)
7069 {
7070 	int i;
7071 
7072 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7073 	root_group->entity.parent = NULL;
7074 	root_group->my_entity = NULL;
7075 	root_group->bfqd = bfqd;
7076 #endif
7077 	root_group->rq_pos_tree = RB_ROOT;
7078 	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7079 		root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7080 	root_group->sched_data.bfq_class_idle_last_service = jiffies;
7081 }
7082 
bfq_init_queue(struct request_queue * q,struct elevator_type * e)7083 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7084 {
7085 	struct bfq_data *bfqd;
7086 	struct elevator_queue *eq;
7087 
7088 	eq = elevator_alloc(q, e);
7089 	if (!eq)
7090 		return -ENOMEM;
7091 
7092 	bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7093 	if (!bfqd) {
7094 		kobject_put(&eq->kobj);
7095 		return -ENOMEM;
7096 	}
7097 	eq->elevator_data = bfqd;
7098 
7099 	spin_lock_irq(&q->queue_lock);
7100 	q->elevator = eq;
7101 	spin_unlock_irq(&q->queue_lock);
7102 
7103 	/*
7104 	 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7105 	 * Grab a permanent reference to it, so that the normal code flow
7106 	 * will not attempt to free it.
7107 	 */
7108 	bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
7109 	bfqd->oom_bfqq.ref++;
7110 	bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7111 	bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7112 	bfqd->oom_bfqq.entity.new_weight =
7113 		bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7114 
7115 	/* oom_bfqq does not participate to bursts */
7116 	bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7117 
7118 	/*
7119 	 * Trigger weight initialization, according to ioprio, at the
7120 	 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7121 	 * class won't be changed any more.
7122 	 */
7123 	bfqd->oom_bfqq.entity.prio_changed = 1;
7124 
7125 	bfqd->queue = q;
7126 
7127 	INIT_LIST_HEAD(&bfqd->dispatch);
7128 
7129 	hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7130 		     HRTIMER_MODE_REL);
7131 	bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7132 
7133 	bfqd->queue_weights_tree = RB_ROOT_CACHED;
7134 	bfqd->num_groups_with_pending_reqs = 0;
7135 
7136 	INIT_LIST_HEAD(&bfqd->active_list);
7137 	INIT_LIST_HEAD(&bfqd->idle_list);
7138 	INIT_HLIST_HEAD(&bfqd->burst_list);
7139 
7140 	bfqd->hw_tag = -1;
7141 	bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7142 
7143 	bfqd->bfq_max_budget = bfq_default_max_budget;
7144 
7145 	bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7146 	bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7147 	bfqd->bfq_back_max = bfq_back_max;
7148 	bfqd->bfq_back_penalty = bfq_back_penalty;
7149 	bfqd->bfq_slice_idle = bfq_slice_idle;
7150 	bfqd->bfq_timeout = bfq_timeout;
7151 
7152 	bfqd->bfq_large_burst_thresh = 8;
7153 	bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7154 
7155 	bfqd->low_latency = true;
7156 
7157 	/*
7158 	 * Trade-off between responsiveness and fairness.
7159 	 */
7160 	bfqd->bfq_wr_coeff = 30;
7161 	bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7162 	bfqd->bfq_wr_max_time = 0;
7163 	bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7164 	bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7165 	bfqd->bfq_wr_max_softrt_rate = 7000; /*
7166 					      * Approximate rate required
7167 					      * to playback or record a
7168 					      * high-definition compressed
7169 					      * video.
7170 					      */
7171 	bfqd->wr_busy_queues = 0;
7172 
7173 	/*
7174 	 * Begin by assuming, optimistically, that the device peak
7175 	 * rate is equal to 2/3 of the highest reference rate.
7176 	 */
7177 	bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7178 		ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7179 	bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
7180 
7181 	spin_lock_init(&bfqd->lock);
7182 
7183 	/*
7184 	 * The invocation of the next bfq_create_group_hierarchy
7185 	 * function is the head of a chain of function calls
7186 	 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7187 	 * blk_mq_freeze_queue) that may lead to the invocation of the
7188 	 * has_work hook function. For this reason,
7189 	 * bfq_create_group_hierarchy is invoked only after all
7190 	 * scheduler data has been initialized, apart from the fields
7191 	 * that can be initialized only after invoking
7192 	 * bfq_create_group_hierarchy. This, in particular, enables
7193 	 * has_work to correctly return false. Of course, to avoid
7194 	 * other inconsistencies, the blk-mq stack must then refrain
7195 	 * from invoking further scheduler hooks before this init
7196 	 * function is finished.
7197 	 */
7198 	bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7199 	if (!bfqd->root_group)
7200 		goto out_free;
7201 	bfq_init_root_group(bfqd->root_group, bfqd);
7202 	bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7203 
7204 	/* We dispatch from request queue wide instead of hw queue */
7205 	blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
7206 
7207 	wbt_disable_default(q);
7208 	blk_stat_enable_accounting(q);
7209 
7210 	return 0;
7211 
7212 out_free:
7213 	kfree(bfqd);
7214 	kobject_put(&eq->kobj);
7215 	return -ENOMEM;
7216 }
7217 
bfq_slab_kill(void)7218 static void bfq_slab_kill(void)
7219 {
7220 	kmem_cache_destroy(bfq_pool);
7221 }
7222 
bfq_slab_setup(void)7223 static int __init bfq_slab_setup(void)
7224 {
7225 	bfq_pool = KMEM_CACHE(bfq_queue, 0);
7226 	if (!bfq_pool)
7227 		return -ENOMEM;
7228 	return 0;
7229 }
7230 
bfq_var_show(unsigned int var,char * page)7231 static ssize_t bfq_var_show(unsigned int var, char *page)
7232 {
7233 	return sprintf(page, "%u\n", var);
7234 }
7235 
bfq_var_store(unsigned long * var,const char * page)7236 static int bfq_var_store(unsigned long *var, const char *page)
7237 {
7238 	unsigned long new_val;
7239 	int ret = kstrtoul(page, 10, &new_val);
7240 
7241 	if (ret)
7242 		return ret;
7243 	*var = new_val;
7244 	return 0;
7245 }
7246 
7247 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)				\
7248 static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
7249 {									\
7250 	struct bfq_data *bfqd = e->elevator_data;			\
7251 	u64 __data = __VAR;						\
7252 	if (__CONV == 1)						\
7253 		__data = jiffies_to_msecs(__data);			\
7254 	else if (__CONV == 2)						\
7255 		__data = div_u64(__data, NSEC_PER_MSEC);		\
7256 	return bfq_var_show(__data, (page));				\
7257 }
7258 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7259 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7260 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7261 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7262 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7263 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7264 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7265 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7266 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7267 #undef SHOW_FUNCTION
7268 
7269 #define USEC_SHOW_FUNCTION(__FUNC, __VAR)				\
7270 static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
7271 {									\
7272 	struct bfq_data *bfqd = e->elevator_data;			\
7273 	u64 __data = __VAR;						\
7274 	__data = div_u64(__data, NSEC_PER_USEC);			\
7275 	return bfq_var_show(__data, (page));				\
7276 }
7277 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7278 #undef USEC_SHOW_FUNCTION
7279 
7280 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)			\
7281 static ssize_t								\
7282 __FUNC(struct elevator_queue *e, const char *page, size_t count)	\
7283 {									\
7284 	struct bfq_data *bfqd = e->elevator_data;			\
7285 	unsigned long __data, __min = (MIN), __max = (MAX);		\
7286 	int ret;							\
7287 									\
7288 	ret = bfq_var_store(&__data, (page));				\
7289 	if (ret)							\
7290 		return ret;						\
7291 	if (__data < __min)						\
7292 		__data = __min;						\
7293 	else if (__data > __max)					\
7294 		__data = __max;						\
7295 	if (__CONV == 1)						\
7296 		*(__PTR) = msecs_to_jiffies(__data);			\
7297 	else if (__CONV == 2)						\
7298 		*(__PTR) = (u64)__data * NSEC_PER_MSEC;			\
7299 	else								\
7300 		*(__PTR) = __data;					\
7301 	return count;							\
7302 }
7303 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7304 		INT_MAX, 2);
7305 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7306 		INT_MAX, 2);
7307 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7308 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7309 		INT_MAX, 0);
7310 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7311 #undef STORE_FUNCTION
7312 
7313 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)			\
7314 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7315 {									\
7316 	struct bfq_data *bfqd = e->elevator_data;			\
7317 	unsigned long __data, __min = (MIN), __max = (MAX);		\
7318 	int ret;							\
7319 									\
7320 	ret = bfq_var_store(&__data, (page));				\
7321 	if (ret)							\
7322 		return ret;						\
7323 	if (__data < __min)						\
7324 		__data = __min;						\
7325 	else if (__data > __max)					\
7326 		__data = __max;						\
7327 	*(__PTR) = (u64)__data * NSEC_PER_USEC;				\
7328 	return count;							\
7329 }
7330 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7331 		    UINT_MAX);
7332 #undef USEC_STORE_FUNCTION
7333 
bfq_max_budget_store(struct elevator_queue * e,const char * page,size_t count)7334 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7335 				    const char *page, size_t count)
7336 {
7337 	struct bfq_data *bfqd = e->elevator_data;
7338 	unsigned long __data;
7339 	int ret;
7340 
7341 	ret = bfq_var_store(&__data, (page));
7342 	if (ret)
7343 		return ret;
7344 
7345 	if (__data == 0)
7346 		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7347 	else {
7348 		if (__data > INT_MAX)
7349 			__data = INT_MAX;
7350 		bfqd->bfq_max_budget = __data;
7351 	}
7352 
7353 	bfqd->bfq_user_max_budget = __data;
7354 
7355 	return count;
7356 }
7357 
7358 /*
7359  * Leaving this name to preserve name compatibility with cfq
7360  * parameters, but this timeout is used for both sync and async.
7361  */
bfq_timeout_sync_store(struct elevator_queue * e,const char * page,size_t count)7362 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7363 				      const char *page, size_t count)
7364 {
7365 	struct bfq_data *bfqd = e->elevator_data;
7366 	unsigned long __data;
7367 	int ret;
7368 
7369 	ret = bfq_var_store(&__data, (page));
7370 	if (ret)
7371 		return ret;
7372 
7373 	if (__data < 1)
7374 		__data = 1;
7375 	else if (__data > INT_MAX)
7376 		__data = INT_MAX;
7377 
7378 	bfqd->bfq_timeout = msecs_to_jiffies(__data);
7379 	if (bfqd->bfq_user_max_budget == 0)
7380 		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7381 
7382 	return count;
7383 }
7384 
bfq_strict_guarantees_store(struct elevator_queue * e,const char * page,size_t count)7385 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
7386 				     const char *page, size_t count)
7387 {
7388 	struct bfq_data *bfqd = e->elevator_data;
7389 	unsigned long __data;
7390 	int ret;
7391 
7392 	ret = bfq_var_store(&__data, (page));
7393 	if (ret)
7394 		return ret;
7395 
7396 	if (__data > 1)
7397 		__data = 1;
7398 	if (!bfqd->strict_guarantees && __data == 1
7399 	    && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7400 		bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7401 
7402 	bfqd->strict_guarantees = __data;
7403 
7404 	return count;
7405 }
7406 
bfq_low_latency_store(struct elevator_queue * e,const char * page,size_t count)7407 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7408 				     const char *page, size_t count)
7409 {
7410 	struct bfq_data *bfqd = e->elevator_data;
7411 	unsigned long __data;
7412 	int ret;
7413 
7414 	ret = bfq_var_store(&__data, (page));
7415 	if (ret)
7416 		return ret;
7417 
7418 	if (__data > 1)
7419 		__data = 1;
7420 	if (__data == 0 && bfqd->low_latency != 0)
7421 		bfq_end_wr(bfqd);
7422 	bfqd->low_latency = __data;
7423 
7424 	return count;
7425 }
7426 
7427 #define BFQ_ATTR(name) \
7428 	__ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7429 
7430 static struct elv_fs_entry bfq_attrs[] = {
7431 	BFQ_ATTR(fifo_expire_sync),
7432 	BFQ_ATTR(fifo_expire_async),
7433 	BFQ_ATTR(back_seek_max),
7434 	BFQ_ATTR(back_seek_penalty),
7435 	BFQ_ATTR(slice_idle),
7436 	BFQ_ATTR(slice_idle_us),
7437 	BFQ_ATTR(max_budget),
7438 	BFQ_ATTR(timeout_sync),
7439 	BFQ_ATTR(strict_guarantees),
7440 	BFQ_ATTR(low_latency),
7441 	__ATTR_NULL
7442 };
7443 
7444 static struct elevator_type iosched_bfq_mq = {
7445 	.ops = {
7446 		.limit_depth		= bfq_limit_depth,
7447 		.prepare_request	= bfq_prepare_request,
7448 		.requeue_request        = bfq_finish_requeue_request,
7449 		.finish_request		= bfq_finish_request,
7450 		.exit_icq		= bfq_exit_icq,
7451 		.insert_requests	= bfq_insert_requests,
7452 		.dispatch_request	= bfq_dispatch_request,
7453 		.next_request		= elv_rb_latter_request,
7454 		.former_request		= elv_rb_former_request,
7455 		.allow_merge		= bfq_allow_bio_merge,
7456 		.bio_merge		= bfq_bio_merge,
7457 		.request_merge		= bfq_request_merge,
7458 		.requests_merged	= bfq_requests_merged,
7459 		.request_merged		= bfq_request_merged,
7460 		.has_work		= bfq_has_work,
7461 		.depth_updated		= bfq_depth_updated,
7462 		.init_hctx		= bfq_init_hctx,
7463 		.init_sched		= bfq_init_queue,
7464 		.exit_sched		= bfq_exit_queue,
7465 	},
7466 
7467 	.icq_size =		sizeof(struct bfq_io_cq),
7468 	.icq_align =		__alignof__(struct bfq_io_cq),
7469 	.elevator_attrs =	bfq_attrs,
7470 	.elevator_name =	"bfq",
7471 	.elevator_owner =	THIS_MODULE,
7472 };
7473 MODULE_ALIAS("bfq-iosched");
7474 
bfq_init(void)7475 static int __init bfq_init(void)
7476 {
7477 	int ret;
7478 
7479 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7480 	ret = blkcg_policy_register(&blkcg_policy_bfq);
7481 	if (ret)
7482 		return ret;
7483 #endif
7484 
7485 	ret = -ENOMEM;
7486 	if (bfq_slab_setup())
7487 		goto err_pol_unreg;
7488 
7489 	/*
7490 	 * Times to load large popular applications for the typical
7491 	 * systems installed on the reference devices (see the
7492 	 * comments before the definition of the next
7493 	 * array). Actually, we use slightly lower values, as the
7494 	 * estimated peak rate tends to be smaller than the actual
7495 	 * peak rate.  The reason for this last fact is that estimates
7496 	 * are computed over much shorter time intervals than the long
7497 	 * intervals typically used for benchmarking. Why? First, to
7498 	 * adapt more quickly to variations. Second, because an I/O
7499 	 * scheduler cannot rely on a peak-rate-evaluation workload to
7500 	 * be run for a long time.
7501 	 */
7502 	ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7503 	ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7504 
7505 	ret = elv_register(&iosched_bfq_mq);
7506 	if (ret)
7507 		goto slab_kill;
7508 
7509 	return 0;
7510 
7511 slab_kill:
7512 	bfq_slab_kill();
7513 err_pol_unreg:
7514 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7515 	blkcg_policy_unregister(&blkcg_policy_bfq);
7516 #endif
7517 	return ret;
7518 }
7519 
bfq_exit(void)7520 static void __exit bfq_exit(void)
7521 {
7522 	elv_unregister(&iosched_bfq_mq);
7523 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7524 	blkcg_policy_unregister(&blkcg_policy_bfq);
7525 #endif
7526 	bfq_slab_kill();
7527 }
7528 
7529 module_init(bfq_init);
7530 module_exit(bfq_exit);
7531 
7532 MODULE_AUTHOR("Paolo Valente");
7533 MODULE_LICENSE("GPL");
7534 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");
7535